HK40056713B - Face image similarity calculation method, apparatus, and device, and storage medium - Google Patents

Description

Methods, apparatus, devices, and storage media for calculating facial image similarity

Technical Field

This invention relates to the field of image processing, and more particularly to a method, apparatus, device, and storage medium for calculating the similarity of human face images.

Background Technology

Multi-frame single-face tracking typically involves first locating the face and then comparing it. This usually spans multiple frames and requires establishing relationships between them. Past techniques have generally progressed from coarse to fine granularity, essentially involving: finding a fixed object range using bounding boxes, then identifying features and comparing them; using keypoints to narrow down the local area (dimensionality reduction through coordinates), and applying Kalman filtering based on the keypoint positions. However, keypoints themselves are inaccurate, and Kalman filtering adds further errors; and finally, detailed segmentation methods (with high annotation costs) to achieve even smaller, more granular comparisons.

These methods all gradually and precisely lock the comparison range to make the comparison objects as free of impurities as possible, thereby improving accuracy. However, the locking range and comparison degree of both methods cannot simultaneously take into account the standard difficulty, speed, and accuracy. There are many other problems encountered: posture (facial rotation), occlusion (environmental occlusion), lighting (facial reflection of light), and resolution (lower resolution results in greater blurriness), which leads to poor generalization.

Summary of the Invention

The main objective of this invention is to improve image recognition efficiency by extracting and fusing features from facial images and determining the correlation between corresponding features of two images.

The first aspect of the present invention provides a method for calculating the similarity of face images, comprising: acquiring two video images containing faces, inputting the video images into a preset face recognition model for recognition, and outputting the region range of the face in the video images; extracting corresponding first face images and second face images from the two video images according to the region range; inputting the first face images and the second face images into the feature layer of a preset attention detection model to extract image features from the first face images and the second face images, respectively obtaining a first image feature of the first face image and a second image feature of the second face image; calculating convolutional attention on the first image features and the second image features respectively, obtaining a first attention image feature and a second attention image feature; calculating the feature similarity between the first attention image feature and the second attention image feature, and determining the image similarity between the first face image and the second face image based on the feature similarity.

Optionally, in a first implementation of the first aspect of the present invention, before acquiring two video images containing faces, inputting the video images into a preset face recognition model for recognition, and outputting the region of the face in the video images, the method further includes: acquiring multiple sample images containing faces in different application scenarios, and using the sample images as a training sample image set; inputting the training sample image set into the backbone network of a preset initial face recognition model, extracting face features from the sample images in the training sample image set to obtain a feature set, wherein the initial face recognition model includes a backbone network and multiple classification networks; calculating the feature vector loss function value of the feature set to obtain multiple feature vector loss function values; calculating the target loss function value of the initial face recognition model based on the multiple feature vector loss function values; iteratively updating the backbone network based on the target loss function value until the target loss function value converges to obtain the target face recognition model.

Optionally, in a second implementation of the first aspect of the present invention, the first face image and the second face image contain global image information. The step of inputting the first face image and the second face image into the feature layer of a preset attention detection model to extract image features from the first face image and the second face image, respectively obtaining a first image feature of the first face image and a second image feature of the second face image, includes: edge extraction of the first face image and the second face image to obtain a first edge image and a second edge image, wherein the first edge image and the second edge image contain edge image information; fusing the global image information and the edge image information to obtain regions in the first face image and the second face image that include the target object; feature extraction of the regions to obtain a first global feature and a first edge feature corresponding to the first face image, and a second global feature and a second edge feature corresponding to the second edge image; feature fusion of the first global feature and the first edge feature to obtain a first image feature of the first face image, and feature fusion of the second global feature and the second edge feature to obtain a second image feature of the second face image.

Optionally, in a third implementation of the first aspect of the present invention, fusing the global image information and the edge image information to obtain the regions containing the target object in the first face image and the second face image includes: extracting features from the edge image information contained in the first edge image and the second edge image through a preset dual-path feature extraction network; and summing the feature extraction results to obtain the region image features containing the target object in the first face image and the second face image.

Optionally, in a fourth implementation of the first aspect of the present invention, the step of calculating convolutional attention on the first image feature and the second image feature respectively to obtain the first attention image feature and the second attention image feature includes: calculating channel attention on the first image feature and the second image feature respectively to obtain a channel attention map of the image feature; calculating spatial attention on the enhanced image feature obtained by merging the image feature and the channel attention map based on the attention mechanism to obtain a spatial attention map of the image feature; and merging the spatial attention map and the enhanced image feature to obtain the first attention image feature of the first face image and the second attention image feature of the second face image respectively.

Optionally, in a fifth implementation of the first aspect of the present invention, the step of calculating channel attention for the first image feature and the second image feature output by the feature layer to obtain a channel attention map of the image feature includes: performing average pooling and max pooling operations on the first image feature and the second image feature respectively to obtain average pooling features and max pooling features; processing the average pooling features using a pre-constructed multilayer perceptron to obtain average pooling parameters, and processing the max pooling features using the multilayer perceptron to obtain max pooling parameters; inputting the sum of the average pooling parameters and the max pooling parameters into an activation module to obtain a first channel attention map of the first image feature and a second channel attention map of the second image feature.

A second aspect of the present invention provides a device for calculating facial image similarity, comprising: a recognition module, configured to acquire two video images containing a face, input the video images into a preset facial recognition model for recognition, and output the region range of the face in the video images; an extraction module, configured to extract a corresponding first face image and a second face image from the two video images based on the region range; a first feature extraction module, configured to input the first face image and the second face image into the feature layer of a preset attention detection model to extract image features from the first face image and the second face image, respectively obtaining a first image feature of the first face image and a second image feature of the second face image; a first calculation module, configured to perform convolutional attention calculation on the first image feature and the second image feature, respectively, to obtain a first attention image feature and a second attention image feature; and a determination module, configured to calculate the feature similarity between the first attention image feature and the second attention image feature, and determine the image similarity between the first face image and the second face image based on the feature similarity.

Optionally, in a first implementation of the second aspect of the present invention, the facial image similarity calculation device includes: an acquisition module, configured to acquire multiple sample images containing faces in different application scenarios, and use the sample images as a training sample image set; a second feature extraction module, configured to input the training sample image set into the backbone network of a preset initial facial recognition model, and extract facial features from the sample images in the training sample image set to obtain a feature set, wherein the initial facial recognition model includes a backbone network and multiple classification networks; a second calculation module, configured to calculate the feature vector loss function value of the feature set to obtain multiple feature vector loss function values; a third calculation module, configured to calculate the target loss function value of the initial facial recognition model based on the multiple feature vector loss function values; and an update module, configured to iteratively update the backbone network based on the target loss function value until the target loss function value converges to obtain a target facial recognition model.

Optionally, in a second implementation of the second aspect of the present invention, the first feature extraction module includes: an edge extraction unit, configured to extract edges from the first face image and the second face image to obtain a first edge image and a second edge image, wherein the first edge image and the second edge image contain edge image information; a fusion unit, configured to fuse the global image information and the edge image information to obtain regions in the first face image and the second face image that include the target object; a feature extraction unit, configured to extract features from the regions to obtain a first global feature and a first edge feature corresponding to the first face image and a second global feature and a second edge feature corresponding to the second edge image; and a feature fusion unit, configured to fuse the first global feature and the first edge feature to obtain a first image feature of the first face image, and to fuse the second global feature and the second edge feature to obtain a second image feature of the second face image.

Optionally, in a third implementation of the second aspect of the present invention, the fusion unit is specifically used to: extract features from the edge image information contained in the first edge image and the second edge image through a preset dual-path feature extraction network, and extract features from the edge image information contained in the first edge image and the second edge image; sum the feature extraction results to obtain the region image features of the target object in the first face image and the second face image.

Optionally, in a fourth implementation of the second aspect of the present invention, the first calculation module includes: a first calculation unit, configured to calculate channel attention for the first image feature and the second image feature respectively, to obtain a channel attention map of the image feature; a second calculation unit, configured to calculate spatial attention for the enhanced image feature obtained by merging the image feature and the channel attention map based on an attention mechanism, to obtain a spatial attention map of the image feature; and a feature merging unit, configured to merge the spatial attention map and the enhanced image feature to obtain a first attention image feature of the first face image and a second attention image feature of the second face image respectively.

Optionally, in a fifth implementation of the second aspect of the present invention, the second computing unit is specifically configured to: perform average pooling and max pooling operations on the first image feature and the second image feature respectively to obtain average pooling features and max pooling features; process the average pooling features using a pre-constructed multilayer perceptron to obtain average pooling parameters, and process the max pooling features using the multilayer perceptron to obtain max pooling parameters; input the sum of the average pooling parameters and the max pooling parameters into an activation module to obtain a first channel attention map of the first image feature and a second channel attention map of the second image feature.

A third aspect of the present invention provides a device for calculating facial image similarity, comprising: a memory and at least one processor, wherein the memory stores instructions, and the memory and the at least one processor are interconnected via a circuit;

The at least one processor invokes the instructions in the memory to cause the face image similarity calculation device to execute the face image similarity calculation method described above.

A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described method for calculating facial image similarity.

The technical solution provided by this invention involves inputting two video frames into a pre-set face recognition model for identification, outputting a first face image and a second face image corresponding to the video images; inputting the face images into the feature layer of a pre-set attention detection model for image feature extraction, obtaining image features of the face images respectively; performing convolutional attention calculation on the image features, obtaining first attention image features and second attention image features of the face images respectively; calculating the feature similarity between the first attention image features and the second attention image features, determining the image similarity between the first face image and the second face image. This solution improves image recognition efficiency by extracting and fusing features from face images and determining the correlation between corresponding features of the two images.

Attached Figure Description

Figure 1 is a schematic diagram of the first embodiment of the method for calculating facial image similarity of the present invention;

Figure 2 is a schematic diagram of the second embodiment of the method for calculating facial image similarity of the present invention;

Figure 3 is a schematic diagram of the third embodiment of the method for calculating facial image similarity of the present invention;

Figure 4 is a schematic diagram of the fourth embodiment of the method for calculating facial image similarity of the present invention;

Figure 5 is a schematic diagram of the fifth embodiment of the method for calculating facial image similarity of the present invention;

Figure 6 is a schematic diagram of the first embodiment of the face image similarity calculation device of the present invention;

Figure 7 is a schematic diagram of a second embodiment of the face image similarity calculation device of the present invention;

Figure 8 is a schematic diagram of an embodiment of the face image similarity calculation device of the present invention.

Detailed Implementation

This invention provides a method, apparatus, device, and storage medium for calculating facial image similarity. The technical solution involves first inputting two video frames into a preset facial recognition model for identification, outputting a first and a second facial image corresponding to the video frames. The facial images are then input into the feature layer of a preset attention detection model for image feature extraction, yielding image features for each facial image. Convolutional attention calculation is performed on these image features to obtain first and second attention image features for each facial image. Finally, the feature similarity between the first and second attention image features is calculated and determined as the image similarity between the first and second facial images. This solution improves image recognition efficiency by extracting and fusing features from facial images and determining the correlation between corresponding features of the two images.

The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to Figure 1. The first embodiment of the method for calculating the similarity of face images in the embodiments of the present invention includes:

101. Acquire two video images containing human faces, input the video images into a preset face recognition model for recognition, and output the region range of the human face in the video images;

In this embodiment, after training the face recognition model, two video images containing faces are obtained from a pre-set database. These video images contain the facial information to be recognized. The video images are then input into the face recognition model.

Face recognition models can identify faces in video images by using bounding boxes for noses, eyes, or other facial features, thus obtaining the region range of each face in the video image.

102. Based on the region range, extract the corresponding first face image and second face image from the two video frames;

In this embodiment, based on the region range, the region range of each face in the video image is then cropped from the second image, thereby extracting the face image corresponding to each video image, namely the first face image and the second face image.

103. Input the first face image and the second face image into the feature layer of the preset attention detection model to extract image features from the first face image and the second face image, respectively, to obtain the first image feature of the first face image and the second image feature of the second face image;

In this embodiment, image feature extraction refers to the computer's ability to recognize numbers, not images. To enable computers to "understand" images and thus possess true "vision," this chapter will explore how to extract useful data or information from images, obtaining "non-image" representations or descriptions, such as numerical values, vectors, and symbols. This process is feature extraction, and the extracted "non-image" representations or descriptions are called features.

In this context, a feature is a corresponding (essential) characteristic or property that distinguishes one class of objects from other classes, or a set of such characteristics and properties. Features are data that can be extracted through measurement or processing. For images, each image possesses its own unique features that distinguish it from other images. Some are natural features that can be intuitively perceived, such as brightness, edges, texture, and color; others require transformation or processing, such as moments, histograms, and principal components. For example, we often combine multiple characteristics of a class of objects to form a feature vector representing that class. If there is only a single numerical feature, the feature vector is a one-dimensional vector; if it is a combination of n characteristics, it is an n-dimensional feature vector. This type of feature vector is often used as input to a recognition system. In fact, an n-dimensional feature is a point located in n-dimensional space, and the task of recognition and classification is to find a partition of this n-dimensional space.

104. Perform convolutional attention calculations on the first image features and the second image features respectively to obtain the first attention image features and the second attention image features;

In this embodiment, the attention detection model includes multiple feature layers connected in sequence; the input of the first feature layer is the input feature, and the input of each feature layer other than the first feature layer is the image feature output by the previous feature layer; the value of the target element in the attention image feature is greater than the value of the target element in the corresponding image feature; the target element refers to the element calculated based on the pixels of the target object in the image to be detected.

105. Calculate the feature similarity between the first attention image features and the second attention image features, and determine the image similarity between the first face image and the second face image based on the feature similarity.

In this embodiment, after obtaining the first attention image features, AND-OR logical operations can be used instead of floating-point operations to calculate the feature similarity between the first attention image features and the second attention image features. This feature similarity can be regarded as the image similarity between the first face image and the second face image.

Specifically, the second face image is the image to be identified, and the first face image is the target image with an identification label. After determining the feature similarity as the image similarity between the first and second face images, if the image similarity is greater than a preset threshold, the identification label is used as the identification result of the second face image. This improves the accuracy and speed of image recognition. The identification label can specifically be a person's identity, classification information, or other identification tags.

In this embodiment of the invention, two video frames are input into a pre-set face recognition model for identification, outputting a first face image and a second face image corresponding to the video images. The face images are then input into the feature layer of a pre-set attention detection model for image feature extraction, yielding image features for each face image. Convolutional attention calculation is performed on these image features to obtain first attention image features and second attention image features for each face image. The feature similarity between the first and second attention image features is calculated and determined as the image similarity between the first and second face images. This scheme improves image recognition efficiency by extracting and fusing features from face images and determining the correlation between corresponding features of the two images.

Please refer to Figure 2. A second embodiment of the method for calculating facial image similarity in this invention includes:

201. Obtain multiple sample images containing human faces from different application scenarios, and use these sample images as a training sample image set;

In this embodiment, one training dataset corresponds to one application scenario, such as a facial recognition scenario and a natural scene scenario. The training dataset can be facial data of different dimensions, open source data, and private data, such as facial data of natural scenes, facial data of Asians, attendance data, facial recognition data, and competition data. The server can extract multiple sample images containing faces from different application scenarios from a pre-set database, preprocess the sample images containing faces, and obtain a preprocessed training data image set.

202. Input the training sample image set into the backbone network of the preset initial face recognition model, and extract face features from the sample images in the training sample image set to obtain a feature set. The initial face recognition model includes a backbone network and multiple classification networks.

In this embodiment, the pre-set initial face recognition model includes a backbone network and multiple classification networks. The output of the backbone network serves as the input to the multiple classification networks. The data processed by the backbone network is classified by the multiple classification networks, thereby enabling face recognition training on the training dataset. The backbone network can be a single convolutional neural network or a combined framework of multiple convolutional neural networks. For example, the backbone network can be the deep residual learning framework ResNet or the object detection network framework ET-YOLOv3, or a combined framework of the deep residual learning framework ResNet and the object detection network framework ET-YOLOv3.

The server can use the backbone network of the initial face recognition model to perform face bounding box recognition, bounding box region segmentation, face key point detection, and face feature vector extraction on each training dataset, obtaining a feature set (i.e., multiple feature sets) corresponding to each training dataset. The convolutional network layers in the backbone network use small convolutional kernels, which retain more features, reduce computation, and improve the efficiency of face feature extraction.

203. Calculate the feature vector loss function value of the feature set to obtain multiple feature vector loss function values;

In this embodiment, a first center vector and a second center vector are calculated, and the distance between each first center vector and the second center vector is calculated. This distance value is used as the feature vector loss function value corresponding to each feature set, thereby obtaining multiple feature vector loss functions. The first center vector is the center vector corresponding to each feature set, or it can be the center vector corresponding to each training data in each feature set. The second center vector can be the second center vector corresponding to all feature sets, or it can be the center vector corresponding to all training data in each feature set.

The server can obtain the number of training data points corresponding to each feature set, and calculate the sum of the first center vectors corresponding to all training data points. The mean of the sums is calculated based on the number of training data points. This mean is the second center vector corresponding to each feature set. The server can also calculate the second center vector using a preset center vector formula.

The server calculates the classification loss function value for each classification dataset using a pre-defined cross-entropy loss function, thus obtaining multiple classification loss function values. This cross-entropy loss function can be a multi-class cross-entropy loss function. Using a multi-class cross-entropy loss function simplifies the differentiation, resulting in faster convergence and quicker updates to the corresponding weight matrix.

204. Based on the loss function values of multiple feature vectors, calculate the target loss function value of the initial face recognition model;

In this embodiment, after obtaining multiple feature vector loss function values and multiple classification loss function values, the number of training datasets is determined. Based on the number of datasets, the average feature vector loss function value and the average classification loss function value are calculated. The sum of the average feature vector loss function value and the average classification loss function value are used as the target loss function value of the face recognition model, or the weighted sum of the average feature vector loss function value and the average classification loss function value is used as the target loss function value of the face recognition model. When each classification network calculates its classification loss function value, the corresponding classification network can be updated in reverse based on the classification loss function value.

205. Iteratively update the backbone network based on the target loss function value until the target loss function value converges to obtain the target face recognition model;

In this embodiment, the network structure and/or weights of the backbone network are iteratively updated based on the target loss function value and a preset number of iterations until the target loss function value converges (i.e., the training accuracy of the face recognition model meets the preset conditions), resulting in an updated face recognition model. The backbone network structure can be updated by adding or removing network layers, by adding other network frameworks, or by modifying the convolutional kernel size and stride. During the iterative update of the backbone network, the server can also optimize the face recognition model using optimization algorithms.

206. Acquire two video images containing human faces, input the video images into a preset face recognition model for recognition, and output the region range of the human face in the video images;

207. Based on the region range, extract the corresponding first face image and second face image from two video frames;

208. Input the first face image and the second face image into the feature layer of the preset attention detection model to extract image features from the first face image and the second face image, respectively, to obtain the first image feature of the first face image and the second image feature of the second face image;

209. Perform convolutional attention calculations on the first image features and the second image features respectively to obtain the first attention image features and the second attention image features;

210. Calculate the feature similarity between the first attention image features and the second attention image features, and determine the image similarity between the first face image and the second face image based on the feature similarity.

Steps 206-210 in this embodiment are similar to steps 101-105 in the first embodiment, and will not be described again here.

In this embodiment of the invention, two video frames are input into a pre-set face recognition model for identification, outputting a first face image and a second face image corresponding to the video images. The face images are then input into the feature layer of a pre-set attention detection model for image feature extraction, yielding image features for each face image. Convolutional attention calculation is performed on these image features to obtain first attention image features and second attention image features for each face image. The feature similarity between the first and second attention image features is calculated and determined as the image similarity between the first and second face images. This scheme improves image recognition efficiency by extracting and fusing features from face images and determining the correlation between corresponding features of the two images.

Please refer to Figure 3. A third embodiment of the method for calculating facial image similarity in this invention includes:

301. Acquire two video images containing human faces, input the video images into a preset face recognition model for recognition, and output the region range of the human face in the video images;

302. Based on the region range, extract the corresponding first face image and second face image from the two video frames;

303. Perform edge extraction on the first face image and the second face image to obtain a first edge image and a second edge image, wherein the first edge image and the second edge image contain edge image information;

In this embodiment, the first face image and the second face image are images to be extracted. The first face image and the second face image can be RGB images (i.e., images composed of the three primary colors red, yellow, and blue), and their formats can be jpg, jpeg, TIFF, PNG, BMP, or PSD, etc., without limitation in this embodiment. The first face image and the second face image include target objects, and the number of target objects can be one or more (i.e., at least two). Additionally, the edge image can be understood as an image used to highlight the boundary between the target object and the background in the first face image and the second face image, as well as the outline of the target object. The target object included in the first face image, the second face image, and the edge image is the same target object, but the representation of the target object in the first face image, the second face image, and the edge image differs.

304. The edge image information contained in the first edge image and the second edge image is extracted by a pre-set dual-path feature extraction network.

In this embodiment, the corresponding pixels in the edge image are obtained by weighted averaging of pixels in a small region of the first and second face images. The dimensions of the first and second face images can be H×W×3; where H represents the height of the first and second face images (e.g., 600), W represents the width of the first and second face images (e.g., 600), and 3 represents the number of the three primary color channels of the first and second face images. The size of the preset convolution kernel can be 3*3, 5*5, or other sizes, which are not limited in this embodiment. For example, if the size of the preset convolution kernel is 3*3, the weight of each unit in the preset convolution kernel can be as follows:

-1 -2 -1 -2 12 -2 -1 -2 -1

Specifically, the method of extracting the edge images corresponding to the first and second face images by performing gradient calculations on the first and second face images according to the preset convolution kernel can be as follows:

A preset convolution kernel Sx is convolved with the first face image and the second face image to obtain [the kernel]. The preset convolution kernel Sx is then transposed to obtain the transposed convolution kernel, which is then convolved with [the kernel] to obtain [the kernel]. By combining [the kernel] and [the kernel], the gradient vectors corresponding to the first face image and the second face image are obtained, including the gradient direction θ and the gradient magnitude. The edge images corresponding to the first face image and the second face image are determined based on the gradient vectors. The edge images include image frequencies representing the degree of grayscale change. Furthermore, it should be noted that regions with rapid gradient magnitude changes can be considered edge regions. The gradient direction θ represents the direction of gradient change. Combining the gradient direction θ and the gradient magnitude allows the edge of the target object in the first face image and the second face image to be determined.

305. Sum the feature extraction results to obtain the regions containing the target object in the first face image and the second face image;

In this embodiment, global image information is used to represent the first face image and the second face image as a whole. Edge image information is used to represent the edges and details of the target object in the first face image and the second face image. The fusion result can be represented as a matrix, corresponding to the first face image and the second face image after edge and detail enhancement of the target object.

The reference image features corresponding to the global image information and the edge image information can be summed, and the summed result can be convolved at a second preset frequency to achieve feature fusion of the global image information and the edge image information, so as to obtain the region image features including the target object in the first face image and the second face image.

306. Extract features from the region to obtain the first global feature and the first edge feature corresponding to the first face image, and the second global feature and the second edge feature corresponding to the second edge image;

In this embodiment, global features are used to characterize the target object as a whole, while edge features are used to highlight the target object at its edges and in detail.

307. Perform feature fusion on the first global feature and the first edge feature to obtain the first image feature of the first face image, and perform feature fusion on the second global feature and the second edge feature to obtain the second image feature of the second face image;

In this embodiment, the output form of the image features can be a matrix. Feature fusion of the aforementioned global and edge features includes: concatenating the global and edge features to obtain a first reference feature, the dimension of which is the sum of the dimensions of the global and edge features; for example, if the dimension of the global feature is 2048 and the dimension of the edge feature is 2048, then the dimension of the first reference feature is 4096; performing dimensionality reduction feature transformation on the first reference feature to obtain a second reference feature, which serves as the image feature corresponding to the target object.

308. Perform convolutional attention calculations on the first image features and the second image features respectively to obtain the first attention image features and the second attention image features;

309. Calculate the feature similarity between the first attention image features and the second attention image features, and determine the image similarity between the first face image and the second face image based on the feature similarity.

Steps 301-302 and 308-309 in this embodiment are similar to steps 101-102 and 104-105 in the first embodiment, and will not be described again here.

In this embodiment of the invention, two video frames are input into a pre-set face recognition model for identification, outputting a first face image and a second face image corresponding to the video images. The face images are then input into the feature layer of a pre-set attention detection model for image feature extraction, yielding image features for each face image. Convolutional attention calculation is performed on these image features to obtain first attention image features and second attention image features for each face image. The feature similarity between the first and second attention image features is calculated and determined as the image similarity between the first and second face images. This scheme improves image recognition efficiency by extracting and fusing features from face images and determining the correlation between corresponding features of the two images.

Please refer to Figure 4. The fourth embodiment of the method for calculating facial image similarity in this invention includes:

401. Acquire two video images containing human faces, input the video images into a preset face recognition model for recognition, and output the region range of the human face in the video images;

402. Based on the region range, extract the corresponding first face image and second face image from the two video frames;

403. Input the first face image and the second face image into the feature layer of the preset attention detection model to extract image features from the first face image and the second face image, respectively, to obtain the first image feature of the first face image and the second image feature of the second face image;

404. Calculate channel attention for the first image feature and the second image feature respectively to obtain the channel attention map of the image features;

In this embodiment, average pooling and max pooling operations are performed on the image features to obtain average pooling features and max pooling features respectively; the average pooling features are processed using a pre-built multilayer perceptron to obtain average pooling parameters, and the max pooling features are processed using a multilayer perceptron to obtain max pooling parameters.

The sum of the average pooling parameters and the max pooling parameters is input into the activation module to obtain the channel attention map of the image features.

The average pooling operation on image features involves moving a pooling window of a pre-defined size (e.g., a 2×2 pooling window) across each feature matrix contained in the image feature. After each move, the area covered by the pooling window is adjacent to the area covered by the pooling window before the move (i.e., one edge of the two areas before and after the move coincides, but the two areas do not overlap). Whenever the pooling window covers a new area, the arithmetic mean of the elements currently covered by the pooling window (in the case of the 2×2 pooling window mentioned above, it can cover 4 elements at a time, i.e., two rows and two columns) is calculated. The calculated result is used as an element in the final average pooling feature. After each element in the image feature has undergone the above average calculation, the average pooling operation on this image feature is completed. All the calculated average values are combined according to the position of the pooling window during the calculation to obtain the average pooling feature corresponding to this image feature.

The process of performing max pooling on image features is basically the same as the average pooling process described above. The difference is that whenever the pooling window covers a new region, the largest element is selected from all elements in that region and used as the result of this calculation (unlike the average pooling operation which uses the average value as the result). Similarly, after each element in the image feature has been selected in the above way, the max pooling process of the image feature is completed. All the selected elements are combined according to the position of the pooling window during the selection to obtain the max pooling feature corresponding to this image feature.

405. Based on the attention mechanism, spatial attention is calculated on the enhanced image features obtained by merging image features and channel attention maps to obtain the spatial attention map of image features;

In this embodiment, average pooling and max pooling operations are performed on the enhanced image features to obtain average pooling enhanced features and max pooling enhanced features, respectively. The average pooling enhanced features and max pooling enhanced features are then merged to obtain merged pooling features. The merged pooling features are then convolved using a convolution kernel of a preset size, and the result of the convolution operation is input into the activation module to obtain a spatial attention map of the image features.

Understandably, for any given feature matrix, only those elements calculated based on the pixels of the target object in the image to be detected (i.e., target elements) are valuable for detecting the target object, while other elements interfere with this process. For example, if the target object in the image to be detected is located in the lower left corner, then the elements in the feature matrix calculated based on the pixels in the lower left corner, also located in the lower left corner, are valuable for detecting the target object. Other elements, such as those located above the target object in the feature matrix, will interfere with the detection process.

406. Merge the spatial attention map and the enhanced image features to obtain the first attention image features of the first face image and the second attention image features of the second face image, respectively;

In this embodiment, channel attention maps and spatial attention maps of image features are calculated, and then these maps are merged with the image features to obtain attention-based image features. This attention calculation increases the effectiveness of convolutional neural network feature extraction, significantly improving the average accuracy of object detection.

Optionally, a normalization layer can be set for the first feature layer. In this case, after the first feature layer outputs image features, it is necessary to: use the normalization layer to perform batch normalization operations on the image features output by the first feature layer to obtain normalized image features; correspondingly, the specific function of the attention layer connected to the first feature layer is: to use the attention layer connected to the feature layer to perform convolutional attention calculation on the normalized image features to obtain attention image features.

In this embodiment, batch normalization is an algorithm designed to address changes in data distribution during training, thereby improving network generalization and accelerating training. During network training, parameters are constantly updated. Updates to the parameters of previous layers lead to changes in the input data distribution of the next layer, requiring that layer to adapt to the new data distribution, significantly impacting training speed. Furthermore, the training process of convolutional neural networks involves learning the data distribution; if the data distribution constantly changes, it reduces the network's generalization ability. Batch normalization essentially preprocesses the data, normalizing it before feeding it into the network. This reduces changes in data distribution, greatly improving the network's generalization and training speed.

407. Calculate the feature similarity between the first attention image features and the second attention image features, and determine the image similarity between the first face image and the second face image based on the feature similarity.

Steps 401-403 and 407 in this embodiment are similar to steps 101-103 and 105 in the first embodiment, and will not be repeated here.

In this embodiment of the invention, two video frames are input into a pre-set face recognition model for identification, outputting a first face image and a second face image corresponding to the video images. The face images are then input into the feature layer of a pre-set attention detection model for image feature extraction, yielding image features for each face image. Convolutional attention calculation is performed on these image features to obtain first attention image features and second attention image features for each face image. The feature similarity between the first and second attention image features is calculated and determined as the image similarity between the first and second face images. This scheme improves image recognition efficiency by extracting and fusing features from face images and determining the correlation between corresponding features of the two images.

Please refer to Figure 5. The fifth embodiment of the method for calculating facial image similarity in this invention includes:

501. Acquire two video images containing human faces, input the video images into a preset face recognition model for recognition, and output the region range of the human face in the video images;

502. Based on the region range, extract the corresponding first face image and second face image from the two video frames;

503. Input the first face image and the second face image into the feature layer of the preset attention detection model to extract image features from the first face image and the second face image, respectively, to obtain the first image feature of the first face image and the second image feature of the second face image;

504. Perform average pooling and max pooling operations on the first image features and the second image features respectively to obtain average pooling features and max pooling features;

In this embodiment, the average pooling operation on the image features refers to using a pooling window with a pre-set size (such as a 2×2 pooling window) to move across each feature matrix contained in the image features. After each move, the area covered by the pooling window is adjacent to the area covered by the pooling window before the move (i.e., one edge of the two areas before and after the move coincides, but the two areas do not overlap). Whenever the pooling window covers a new area, the arithmetic mean of the elements currently covered by the pooling window (with the above-mentioned 2×2 pooling window, it can cover 4 elements at a time, i.e., two rows and two columns) is calculated. The calculated result is used as one element in the final average pooling feature. After each element in the image feature has undergone the above average calculation, the average pooling operation on this image feature is completed. All the calculated average values are combined according to the position of the pooling window during the calculation to obtain the average pooling feature corresponding to this image feature.

The process of performing max pooling on image features is basically the same as the average pooling process described above. The difference is that whenever the pooling window covers a new region, the largest element is selected from all elements in that region and used as the result of this calculation (unlike the average pooling operation which uses the average value as the result). Similarly, after each element in the image feature has been selected in the above way, the max pooling process of the image feature is completed. All the selected elements are combined according to the position of the pooling window during the selection to obtain the max pooling feature corresponding to this image feature.

505. Use a pre-built multilayer perceptron to process the average pooling features to obtain the average pooling parameters, and use a multilayer perceptron to process the max pooling features to obtain the max pooling parameters.

In this embodiment, the multilayer perceptron is a feedforward artificial neural network model that maps multiple input datasets to a single output dataset. Specifically, in this scheme, the multilayer perceptron is used to map max-pooling features and average-pooling features to a one-dimensional vector containing C parameters, i.e., a vector of the form (A1, A2, ..., AC-1, AC).

Where C is the number of channels of the image features input to this attention layer (the number of feature matrices contained in an image feature is called the number of channels).

The two one-dimensional vectors output by the multilayer perceptron are the average pooling parameter and the max pooling parameter mentioned in the previous calculation process.

506. Input the sum of the average pooling parameter and the max pooling parameter into the activation module to obtain the first channel attention map of the first image feature and the second channel attention map of the second image feature;

In this embodiment, the channel attention map is obtained by performing activation operations on the two one-dimensional vectors output by the multilayer perceptron using an activation function (equivalent to inputting the sum of the average pooling parameter and the max pooling parameter into the activation module). The channel attention map is also a one-dimensional vector containing C parameters.

In this embodiment, the function of the channel attention map is to highlight meaningful feature matrices within the image features. In an image, the salience of features of each object differs across different feature matrices of the same image feature set. For example, the features of a car might be more prominent on the first feature matrix, while the features of a house might be more prominent on the second. Clearly, in the object detection application scenario, feature matrices containing more prominent features of the target object are meaningful feature matrices, while other feature matrices are meaningless.

507. Based on the attention mechanism, spatial attention is calculated on the enhanced image features obtained by merging image features and channel attention maps to obtain the spatial attention map of image features;

508. Merge the spatial attention map and the enhanced image features to obtain the first attention image features of the first face image and the second attention image features of the second face image, respectively;

509. Calculate the feature similarity between the first attention image features and the second attention image features, and determine the image similarity between the first face image and the second face image based on the feature similarity.

Steps 501-503 and 509 in this embodiment are similar to steps 101-103 and 105 in the first embodiment, and will not be described again here.

In this embodiment of the invention, two video frames are input into a pre-set face recognition model for identification, outputting a first face image and a second face image corresponding to the video images. The face images are then input into the feature layer of a pre-set attention detection model for image feature extraction, yielding image features for each face image. Convolutional attention calculation is performed on these image features to obtain first attention image features and second attention image features for each face image. The feature similarity between the first and second attention image features is calculated and determined as the image similarity between the first and second face images. This scheme improves image recognition efficiency by extracting and fusing features from face images and determining the correlation between corresponding features of the two images.

The method for calculating face image similarity in the embodiments of the present invention has been described above. The following describes the apparatus for calculating face image similarity in the embodiments of the present invention. Referring to Figure 6, the first embodiment of the apparatus for calculating face image similarity in the embodiments of the present invention includes:

The recognition module 601 is used to acquire two video images containing human faces, input the video images into a preset face recognition model for recognition, and output the region range of the human face in the video images;

Extraction module 602 is used to extract the corresponding first face image and second face image from the two video images according to the region range;

The first feature extraction module 603 is used to input the first face image and the second face image into the feature layer of a preset attention detection model to extract image features from the first face image and the second face image, respectively obtaining the first image feature of the first face image and the second image feature of the second face image;

The first calculation module 604 is used to calculate convolutional attention on the first image features and the second image features respectively, to obtain first attention image features and second attention image features;

The determination module 605 is used to calculate the feature similarity between the first attention image feature and the second attention image feature, and to determine the image similarity between the first face image and the second face image based on the feature similarity.

In this embodiment of the invention, two video frames are input into a pre-set face recognition model for identification, outputting a first face image and a second face image corresponding to the video images. The face images are then input into the feature layer of a pre-set attention detection model for image feature extraction, yielding image features for each face image. Convolutional attention calculation is performed on these image features to obtain first attention image features and second attention image features for each face image. The feature similarity between the first and second attention image features is calculated and determined as the image similarity between the first and second face images. This scheme improves image recognition efficiency by extracting and fusing features from face images and determining the correlation between corresponding features of the two images.

Please refer to Figure 7, which illustrates a second embodiment of the face image similarity calculation device according to the present invention. This face image similarity calculation device specifically includes:

The recognition module 601 is used to acquire two video images containing human faces, input the video images into a preset face recognition model for recognition, and output the region range of the human face in the video images;

Extraction module 602 is used to extract the corresponding first face image and second face image from the two video images according to the region range;

The first feature extraction module 603 is used to input the first face image and the second face image into the feature layer of a preset attention detection model to extract image features from the first face image and the second face image, respectively obtaining the first image feature of the first face image and the second image feature of the second face image;

The first calculation module 604 is used to calculate convolutional attention on the first image features and the second image features respectively to obtain first attention image features and second attention image features;

The determination module 605 is used to calculate the feature similarity between the first attention image feature and the second attention image feature, and to determine the image similarity between the first face image and the second face image based on the feature similarity.

In this embodiment, the facial image similarity calculation device includes:

The acquisition module 606 is used to acquire multiple video images containing human faces in different application scenarios, and use the video images as a training sample image set.

The second feature extraction module 607 is used to input the training sample image set into the backbone network of the preset initial face recognition model, and to extract face features from the video images in the training sample image set to obtain a feature set. The initial face recognition model includes a backbone network and multiple classification networks.

The second calculation module 608 is used to calculate the feature vector loss function value of the feature set and obtain multiple feature vector loss function values.

The third calculation module 609 is used to calculate the target loss function value of the initial face recognition model based on the multiple feature vector loss function values;

The update module 610 is used to iteratively update the backbone network according to the target loss function value until the target loss function value converges, thereby obtaining the target face recognition model.

In this embodiment, the first feature extraction module 603 includes:

Edge extraction unit 6031 is used to extract edges from the first face image and the second face image to obtain a first edge image and a second edge image;

The fusion unit 6032 is used to fuse the global image information contained in the first face image and the second face image and the edge image information contained in the first edge image and the second edge image to obtain the region containing the target object in the first face image and the second face image.

Feature extraction unit 6033 is used to extract features from the region to obtain a first global feature and a first edge feature corresponding to the first face image and a second global feature and a second edge feature corresponding to the second edge image;

The feature fusion unit 6034 is used to perform feature fusion on the first global feature and the first edge feature, as well as the second global feature and the second edge feature, respectively, to obtain the first image feature of the first face image and the second image feature of the second face image.

In this embodiment, the fusion unit 6032 is specifically used for:

The edge image information contained in the first edge image and the second edge image is extracted by a pre-set dual-path feature extraction network.

The feature extraction results are summed to obtain the region image features of the target object in the first face image and the second face image.

In this embodiment, the first computing module 604 includes:

The first calculation unit 6041 is used to calculate channel attention for the first image features and the second image features output by the feature layer, respectively, to obtain a channel attention map of the image features;

The second computing unit 6042 is used to perform spatial attention calculation on the enhanced image features obtained by merging the image features and the channel attention map based on the attention mechanism, so as to obtain the spatial attention map of the image features.

The feature merging unit 6043 is used to merge the spatial attention map and the enhanced image features to obtain the first attention image features of the first face image and the second attention image features of the second face image, respectively.

In this embodiment, the second computing unit 6042 is specifically used for:

Average pooling and max pooling operations are performed on the first image features and the second image features respectively to obtain average pooling features and max pooling features;

The average pooling feature is processed using a pre-built multilayer perceptron to obtain the average pooling parameter, and the max pooling feature is processed using the multilayer perceptron to obtain the max pooling parameter.

The sum of the average pooling parameter and the max pooling parameter is input into the activation module to obtain the first channel attention map of the first image feature and the second channel attention map of the second image feature.

In this embodiment of the invention, two video frames are input into a pre-set face recognition model for identification, outputting a first face image and a second face image corresponding to the video images. The face images are then input into the feature layer of a pre-set attention detection model for image feature extraction, yielding image features for each face image. Convolutional attention calculation is performed on these image features to obtain first attention image features and second attention image features for each face image. The feature similarity between the first and second attention image features is calculated and determined as the image similarity between the first and second face images. This scheme improves image recognition efficiency by extracting and fusing features from face images and determining the correlation between corresponding features of the two images.

Figures 6 and 7 above describe in detail the face image similarity calculation device in the embodiments of the present invention from the perspective of modular functional entities. The following describes in detail the face image similarity calculation device in the embodiments of the present invention from the perspective of hardware processing.

Figure 8 is a schematic diagram of a face image similarity calculation device provided in an embodiment of the present invention. This face image similarity calculation device 800 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 810 (e.g., one or more processors) and a memory 820, and one or more storage media 830 (e.g., one or more mass storage devices) storing application programs 833 or data 832. The memory 820 and storage media 830 can be temporary or persistent storage. The program stored in the storage media 830 may include one or more modules (not shown in the figure), each module including a series of instruction operations on the face image similarity calculation device 800. Furthermore, the processor 810 may be configured to communicate with the storage media 830 and execute a series of instruction operations in the storage media 830 on the face image similarity calculation device 800 to implement the steps of the face image similarity calculation method provided in the above-described method embodiments.

The face image similarity computing device 800 may also include one or more power supplies 840, one or more wired or wireless network interfaces 850, one or more input/output interfaces 860, and/or one or more operating systems 831, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that the face image similarity computing device structure shown in FIG8 does not constitute a limitation on the face image similarity computing device provided in this application, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the above-described method for calculating the similarity of facial images.

Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (7)

1. A method for calculating the similarity of facial images, characterized in that the method for calculating the similarity of facial images includes: Two video images containing human faces are obtained from a pre-set database. Based on a pre-set face recognition model, the faces are identified by the facial features in the video images, and the output is the region range of the face in the video images. Based on the area range, the corresponding first face image and second face image are extracted from the two video images, wherein the first face image and the second face image contain global image information; The first face image and the second face image are input into the feature layer of a preset attention detection model to extract image features from the first face image and the second face image, respectively, to obtain the first image feature of the first face image and the second image feature of the second face image; Convolutional attention is calculated on the first image features and the second image features respectively to obtain the first attention image features and the second attention image features; Calculate the feature similarity between the first attention image features and the second attention image features, and determine the image similarity between the first face image and the second face image based on the feature similarity; The step of inputting the first face image and the second face image into the feature layer of a preset attention detection model to extract image features from the first face image and the second face image to obtain the first image feature of the first face image and the second image feature of the second face image includes: performing edge extraction on the first face image and the second face image to obtain a first edge image and a second edge image, wherein the first edge image and the second edge image contain edge image information; performing feature extraction on the edge image information contained in the first edge image and the second edge image through a preset dual-path feature extraction network, and performing feature extraction on the edge image information contained in the first edge image and the second edge image; summing the feature extraction results to obtain the region containing the target object in the first face image and the second face image; performing feature extraction on the region to obtain the first global feature and the first edge feature corresponding to the first face image, and the second global feature and the second edge feature corresponding to the second edge image; performing feature fusion on the first global feature and the first edge feature to obtain the first image feature of the first face image, and performing feature fusion on the second global feature and the second edge feature to obtain the second image feature of the second face image; The step of calculating convolutional attention on the first image feature and the second image feature respectively to obtain the first attention image feature and the second attention image feature includes: calculating channel attention on the first image feature and the second image feature respectively to obtain the channel attention map of the image feature; calculating spatial attention on the enhanced image feature obtained by merging the image feature and the channel attention map based on the attention mechanism to obtain the spatial attention map of the image feature; and merging the spatial attention map and the enhanced image feature to obtain the first attention image feature of the first face image and the second attention image feature of the second face image respectively. 2. The method for calculating facial image similarity according to claim 1, characterized in that, before obtaining two video images containing faces from a preset database, identifying faces based on facial features in the video images using a preset facial recognition model, and outputting the region range of the face in the video images, the method further includes: Acquire multiple sample images containing human faces from different application scenarios, and use these sample images as a training sample image set; The training sample image set is input into the backbone network of the preset initial face recognition model, and face features are extracted from the sample images in the training sample image set to obtain a feature set. The initial face recognition model includes a backbone network and multiple classification networks. Calculate the feature vector loss function value of the feature set to obtain multiple feature vector loss function values; Calculate the target loss function value of the initial face recognition model based on the multiple feature vector loss function values; The backbone network is iteratively updated based on the target loss function value until the target loss function value converges, thus obtaining the target face recognition model. 3. The method for calculating facial image similarity according to claim 1, characterized in that, the step of calculating channel attention for the first image features and the second image features output by the feature layer respectively to obtain the channel attention map of the image features includes: Average pooling and max pooling operations are performed on the first image features and the second image features respectively to obtain average pooling features and max pooling features; The average pooling feature is processed using a pre-built multilayer perceptron to obtain the average pooling parameter, and the max pooling feature is processed using the multilayer perceptron to obtain the max pooling parameter. The sum of the average pooling parameter and the max pooling parameter is input into the activation module to obtain the first channel attention map of the first image feature and the second channel attention map of the second image feature. 4. A device for calculating the similarity of facial images, characterized in that the device comprises: The recognition module is used to obtain two video images containing human faces from a preset database, identify the human face based on the facial features in the video images using a preset face recognition model, and output the region range of the human face in the video images. The extraction module is used to extract a corresponding first face image and a second face image from the two video images according to the region range, wherein the first face image and the second face image contain global image information; The first feature extraction module is used to input the first face image and the second face image into the feature layer of a preset attention detection model to extract image features from the first face image and the second face image, respectively obtaining the first image feature of the first face image and the second image feature of the second face image; The first calculation module is used to calculate convolutional attention on the first image features and the second image features respectively, to obtain first attention image features and second attention image features; The determination module is used to calculate the feature similarity between the first attention image features and the second attention image features, and to determine the image similarity between the first face image and the second face image based on the feature similarity. The step of inputting the first face image and the second face image into the feature layer of a preset attention detection model to extract image features from the first face image and the second face image to obtain the first image feature of the first face image and the second image feature of the second face image includes: performing edge extraction on the first face image and the second face image to obtain a first edge image and a second edge image, wherein the first edge image and the second edge image contain edge image information; performing feature extraction on the edge image information contained in the first edge image and the second edge image through a preset dual-path feature extraction network, and performing feature extraction on the edge image information contained in the first edge image and the second edge image; summing the feature extraction results to obtain the region containing the target object in the first face image and the second face image; performing feature extraction on the region to obtain the first global feature and the first edge feature corresponding to the first face image, and the second global feature and the second edge feature corresponding to the second edge image; performing feature fusion on the first global feature and the first edge feature to obtain the first image feature of the first face image, and performing feature fusion on the second global feature and the second edge feature to obtain the second image feature of the second face image; The step of calculating convolutional attention on the first image feature and the second image feature respectively to obtain the first attention image feature and the second attention image feature includes: calculating channel attention on the first image feature and the second image feature respectively to obtain the channel attention map of the image feature; calculating spatial attention on the enhanced image feature obtained by merging the image feature and the channel attention map based on the attention mechanism to obtain the spatial attention map of the image feature; and merging the spatial attention map and the enhanced image feature to obtain the first attention image feature of the first face image and the second attention image feature of the second face image respectively. 5. The facial image similarity calculation device according to claim 4, characterized in that the facial image similarity calculation device further comprises: The acquisition module is used to acquire multiple sample images containing human faces in different application scenarios, and use the sample images as a training sample image set; The second feature extraction module is used to input the training sample image set into the backbone network of the preset initial face recognition model, and to extract face features from the sample images in the training sample image set to obtain a feature set. The initial face recognition model includes a backbone network and multiple classification networks. The second calculation module is used to calculate the feature vector loss function value of the feature set and obtain multiple feature vector loss function values. The third calculation module is used to calculate the target loss function value of the initial face recognition model based on the multiple feature vector loss function values; The update module is used to iteratively update the backbone network based on the target loss function value until the target loss function value converges, thereby obtaining the target face recognition model. 6. A facial image similarity calculation device, characterized in that the facial image similarity calculation device comprises: a memory and at least one processor, wherein the memory stores instructions, and the memory and the at least one processor are interconnected via a circuit; The at least one processor invokes the instructions in the memory to cause the face image similarity calculation device to perform the steps of the face image similarity calculation method as described in any one of claims 1-3. 7. A computer-readable storage medium storing a computer program, characterized in that, when executed by a processor, the computer program implements the steps of the method for calculating the similarity of face images as described in any one of claims 1-3.