CN109815814A - A face detection method based on convolutional neural network - Google Patents

Abstract

The invention discloses a face detection method based on a convolutional neural network, comprising the steps of: step (1), establishing a database; step (2), performing an image in the database; preprocessing; Build a good degree learning network; step (4), test the training results, and accurately detect the faces that are occluded, different angles, and sides in the picture, as well as the smaller and more blurred faces in the picture. The rate is higher, the network structure is simple, the iteration parameters are less, and the training time is shorter.

Description

A face detection method based on convolutional neural network

technical field

The invention belongs to the fields of computer vision and artificial intelligence, and particularly relates to a face detection method based on a convolutional neural network.

Background technique

Face detection refers to the process of determining the position and size of a face in an image of a face. It is an important part in the field of computer vision and a key step in preprocessing for face recognition. Its detection accuracy is very high. It also determines the accuracy of face recognition to a large extent and has a great impact on subsequent work. Therefore, the research on face detection has great significance and practical value.

Face detection has a wide range of applications in real life: such as identity authentication and security protection, in media and entertainment, mobile phones, digital cameras and other electronic products, as well as image retrieval and so on. Face detection methods can be roughly divided into traditional detection methods (including detection methods based on matching templates, distance-based detection methods, etc.) and detection methods based on deep learning.

In recent years, deep learning has been continuously improved and developed, and it has been widely used in both classification and regression tasks. Face detection methods based on deep learning are also developing continuously, but for the current methods, taking the most commonly used MTCNN method as an example, its recognition speed is not fast enough, and the recognition accuracy is not high enough, especially for images and videos. Occlusions, or different angles, sides, and smaller faces in the frame are not easy to detect. As a preprocessing step in the face recognition process, the accuracy of face detection also greatly affects the accuracy of subsequent recognition work, so it is very important to solve these problems.

SUMMARY OF THE INVENTION

Based on the prior art, the present invention proposes a face detection method based on a convolutional neural network, in particular, it relates to the detection of faces in the picture due to illumination, occlusion, or in the side state and very small in the picture, By establishing a new database, building a convolutional neural network, and by adjusting the hyperparameters and iteratively training the network, a better detection effect can be obtained, thereby effectively detecting faces.

The present invention proposes a face detection method based on a convolutional neural network, the method comprising the following steps:

A face detection method based on convolutional neural network, the method includes the following steps:

Step 1. Establish a database to obtain image data for preprocessing to construct a convolutional neural network;

Step 2, performing four iterative operations on the preprocessed data through the image feature analysis module in the convolutional neural network to generate image feature parameters;

Step 3. Generate an image one-dimensional vector by operating on the image feature parameters through the fully connected layer in the convolutional neural network;

Step 4: Classify and regress the one-dimensional vector of the image through the classification layer in the convolutional neural network to obtain the position coordinates of the face image.

The process of preprocessing data by the image feature analysis module in the step 2 includes the following steps:

The convolution layer of the image feature analysis module described in step 2.1 extracts image features by convolving the weights and parameters of the preprocessed data;

Step 2.2, the activation function layer of the image feature analysis module uses the ReLu function to perform nonlinear operations on the image features to obtain nonlinear feature map parameters;

Step 2.3: The maximum pooling layer of the image feature analysis module reduces the parameters of the nonlinear feature map.

In the step 4, the classification layer classifies and regresses the one-dimensional vector of the image: including the following steps.

Step 4.1, through the optimization method of the stochastic gradient descent method, iterative weights are performed on the one-dimensional vector of the image, so as to continuously adjust the loss function, and continuously adjust the hyperparameters during training to obtain the best training results. The hyperparameters include: iteration times, batches, maximum number of iterations, learning rate;

Step 4.2, the loss function selected in the classification process is to combine the central loss function with the softmax loss function

method, the specific expression is:

Among them, L _S is the softmax loss function, L _c is the central loss function, and λ is the coefficient, indicating that the weight of the two is λ=0.1 here. In the formula, Wx+b is the output of the fully connected layer, after log, it represents the probability that x _i belongs to the category y _i , and C represents the feature center of the category;

Step 4.3, the loss function used in the regression process is: Euclidean distance loss function, the specific expression is as follows:

y _i ∈R ⁴

in, is the output result of the network prediction, and y is the marked real label, that is, the coordinates of the 68 face key points. Step 4.4, compare the coordinates of the 68 face key points output under the optimal weight condition with the labeled face key point coordinates and faces in the database, so as to calculate the convolutional neural network for face detection. 's accuracy.

beneficial effect

Compared with the prior art, the present invention provides a face detection method based on a convolutional neural network. The detection accuracy of the face is relatively high, and the network structure is simple, the iterative parameters are few, and the training time is short.

Description of drawings

Fig. 1 is a kind of flow chart of the face detection method based on convolutional neural network;

FIG. 2 is a connection mode of a convolutional neural network used in a face recognition method based on a convolutional neural network proposed by the present invention, which includes four convolutional layers, four ReLu activation function layers, and four maximum pooling layers. , two fully connected layers, the last fully connected layer is the softmax classification layer;

Detailed ways

Below in conjunction with accompanying drawing, the present invention is described in further detail:

As shown in Figure 1, it is a flow chart of a face detection method based on a convolutional neural network.

A face detection method based on convolutional neural network, the method includes the following steps:

Step 1 (110) establishes a database to obtain image data for preprocessing to construct a convolutional neural network;

In this step, a database is established to obtain image data, that is, the established database contains pictures with the following requirements: the picture contains at least one face, and the position of the face is not required, preferably a person who is not in the center and is far away face; and the background of the face is complex and diverse, including various indoor and outdoor scenes; the position of the face in the image is marked with a rectangular frame, and the face includes eyebrows, eyes, nose, mouth, and facial contours. 68 key points including. Image clarity is not required. The established database contains a total of 6,000 images containing human faces and labeled.

In this step, the images in the database are preprocessed, and the images in the established database are firstly subjected to the spatial pyramid pooling operation. This operation can obtain multiple images with different pixels and different scales from one image, which is convenient for extracting images from the database. A feature vector of a fixed size is extracted from the multi-scale features; all images generated in the above steps are randomly mirrored, including up and down, left and right mirror operations; 4/5 of the database images processed in the above steps are used as the training database, 1 /5 is used as a test database;

Step 2, (210) by the image feature analysis module in the convolutional neural network, perform four iterative operations on the preprocessed data to generate image feature parameters;

The convolution layer of the image feature analysis module described in step 2.1 extracts image features by convolving the weights and parameters of the preprocessed data;

2.2. The activation function layer of the image feature analysis module uses the ReLu function to perform nonlinear operations on image features to obtain nonlinear feature map parameters;

Step 2.3: The maximum pooling layer of the image feature analysis module reduces the parameters of the nonlinear feature map.

In the present invention, the preprocessed test database image is sent into the trained neural network, the test image is subjected to the trained neural network weight matrix, the features are extracted, and then the classifier is passed to output the result of classification and regression, and the result of the classification In the form of probability, if the probability of being judged as a face is greater than the probability of being judged as a non-face, it is judged as a face, and the part judged as a face will be marked with a rectangular frame; the result of the regression is the 68 key points of the face part in the picture mark with a keypoint, and return the coordinates of the mark.

Step 3. (310) Generate a one-dimensional image vector by operating on the image feature parameters through the fully connected layer in the convolutional neural network.

Step 4. (410) Classify and regress the one-dimensional vector of the image through the classification layer in the convolutional neural network to obtain the position coordinates of the face image. In the step 4, the classification layer performs the classification and regression process on the one-dimensional vector of the image: including the following steps:

Step 4.1, through the optimization method of the stochastic gradient descent method, iterative weights are performed on the one-dimensional vector of the image, so as to continuously adjust the loss function, and continuously adjust the hyperparameters during training to obtain the best training results. The hyperparameters include: iteration times, batches, maximum number of iterations, learning rate;

Step 4.2, the loss function selected in the classification process is a method of combining the central loss function and the softmax loss function, and the specific expression is:

Among them, L _S is the softmax loss function, L _c is the central loss function, and λ is the coefficient, indicating that the weight of the two is λ=0.1 here. In the formula, Wx+b is the output of the fully connected layer, after log, it represents the probability that x _i belongs to the category y _i , and C represents the feature center of the category;

Step 4.3, the loss function used in the regression process is: Euclidean distance loss function, the specific expression is as follows:

y _i ∈R ⁴

in, is the output result of the network prediction, and y is the marked real label, that is, the coordinates of the 68 face key points. Step 4.4, compare the coordinates of the 68 face key points output under the optimal weight condition with the labeled face key point coordinates and faces in the database, so as to calculate the convolutional neural network for face detection. 's accuracy.

The training task of the present invention is divided into two parts as a whole: classification and regression. Classification refers to the problem of face detection as a two-class problem of face and non-face; regression refers to the process of returning the coordinates of the face frame and the coordinates of the 68 key points of the face after neural network training. , so as to complete the purpose of detecting faces. After continuous iterative updating of the weights in the network to reduce the loss function, the optimal weights are finally obtained; the recognition results output under the optimal weights are compared with the labeled face key point coordinates and face in the database. For comparison, the accuracy of this convolutional neural network for face detection is calculated.

As shown in Figure 2, the convolutional neural network used in the face recognition method based on the convolutional neural network proposed by the present invention includes four convolutional layers, four ReLu activation function layers, four maximum pooling layers, two The last fully connected layer is the softmax classification layer. The functions are: the convolution layer is used to extract the features of the image by convolving the weights of the convolution layer with the parameters; the activation function layer is to increase the nonlinear ability of the network, and the ReLu function refers to y=max. (0,x) this function; the maximum pooling layer is to reduce the output size and the amount of parameters; the fully connected layer is to map the extracted features into a one-dimensional vector; the classification layer is to extract the The features are classified into two parts: face and non-face, and the coordinates of 68 key points that return to the face. The whole training process is as follows: first, the parameters in the convolution layer and the fully connected layer are randomly initialized. After the images in the established database are sent to this network, after four convolution, activation, and pooling layers, the face is obtained. features, and then obtain a fixed-size feature vector through the fully connected layer, and finally obtain the coordinates of the location of the face through the classification layer. The classification layer is to classify the face and non-face in the features extracted by the aforementioned network, and to return the coordinates of the 68 key points of the face. In the process of network training, the data propagates forward along the network, and the error obtained through the loss function propagates backward along the network, so that the parameters in the convolutional layer and the fully connected layer are continuously optimized. Get a good training effect.

The whole training process of the present invention is as follows: first, the parameters in the convolution layer and the fully connected layer are randomly initialized, and after the images in the established database are sent to this network, after four convolution, activation, and pooling layers, The features of the face are obtained, and then a fixed-size feature vector is obtained through the fully connected layer, and finally the coordinates of the location of the face are obtained through the classification layer. In the process of network training, the data propagates forward along the network, and the error obtained through the loss function propagates backward along the network, so that the parameters in the convolutional layer and the fully connected layer are continuously optimized. Get a good training effect. This step is to obtain the optimal parameters by training the database. During the whole training process, the error between the actual label and the predicted result is represented by the loss function, that is, by minimizing the loss function, iterative training is performed continuously, and the optimal parameters are obtained when the final loss function reaches the minimum. Among them, the parameters to be trained include the convolution kernel and bias of the convolution layer, and the neuron parameters in the fully connected layer. During the whole training process, the data is propagated forward, the error calculated by the loss function is back propagated, and through the gradient descent method, the network finds the global optimal point in the process of continuous iteration, and the optimal parameters are obtained at this time. After the training is over, that is, after obtaining the optimal network parameters, the optimal parameters are substituted into the entire network. At this time, the entire network has the capability of face detection, and face detection can be performed. Afterwards, the accuracy of this neural network for face detection can be obtained through testing.

The above descriptions are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention are included in the protection of the present invention. within the range.

Claims (3)

1. a face detection method based on convolutional neural network, is characterized in that, this method comprises the following steps: Step 1. Establish a database to obtain image data for preprocessing to construct a convolutional neural network; Step 2, performing four iterative operations on the preprocessed data through the image feature analysis module in the convolutional neural network to generate image feature parameters; Step 3. Generate an image one-dimensional vector by operating on the image feature parameters through the fully connected layer in the convolutional neural network; Step 4: Classify and regress the one-dimensional vector of the image through the classification layer in the convolutional neural network to obtain the position coordinates of the face image. 2. the face detection method based on convolutional neural network as claimed in claim 1, is characterized in that, in described step 2, image feature analysis module comprises the following steps to preprocessing data process: Step 2.1, the convolution layer of the image feature analysis module extracts image features by convolving the weights and parameters of the preprocessed data; Step 2.2, the activation function layer of the image feature analysis module uses the ReLu function to perform nonlinear operations on the image features to obtain nonlinear feature map parameters; Step 2.3: The maximum pooling layer of the image feature analysis module reduces the parameters of the nonlinear feature map. 3. The face detection method based on a convolutional neural network as claimed in claim 1, wherein in the step 4, the classification layer performs a classification and regression process on the image one-dimensional vector: comprising the following steps. Step 4.1. Iterative weights are performed on the one-dimensional vector of the image through the optimization method of the stochastic gradient descent method, so as to continuously adjust the loss function, and continuously adjust the hyperparameters during training to obtain the best training results. The hyperparameters include: iteration times, batches, maximum number of iterations, learning rate; Step 4.2. The loss function selected in the classification process is a method of combining the central loss function and the softmax loss function. The specific expression is: Among them, L _S is the softmax loss function, L _c is the central loss function, and λ is the coefficient, indicating that the weight of the two is λ=0.1 here. In the formula, Wx+b is the output of the fully connected layer, after log, it represents the probability that x _i belongs to the category y _i , and C represents the feature center of the category; Step 4.3. The loss function used in the regression process is: Euclidean distance loss function, the specific expression is as follows: y _i ∈R ⁴ in, is the output result of the network prediction, and y is the marked real label, that is, the coordinates of the 68 face key points. Step 4.4. Compare the coordinates of the 68 face key points output under the optimal weight condition with the labeled face key point coordinates and faces in the database, so as to calculate the convolutional neural network for face detection. 's accuracy.