TWI843591B - Method for creating flaw image detection model, method for detecting flaw image and electronic device - Google Patents

Abstract

A method for creating a flaw image detection model includes: generating a first training image by using a flawless sample image and a flawed sample image; inputting the first training image to an autoencoder to output a reconstructed image and calculating a first reconstruction error; computing a difference between the first training image and the reconstructed image to generate a reconstruction error image; generating a second training image by using the first training image, the reconstructed image, and the reconstruction error image; inputting the second training image to an image segmentation network to generate a mask image and calculate a second reconstruction error; calculating a total reconstruction error based on the first reconstruction error and the second reconstruction error; optimizing parameters of the autoencoder and the image segmentation network by the total reconstruction error; and repeating above steps until the flaw image detection model is established and the parameters are optimized.

Description

Method for establishing image defect detection model, method for detecting defective image and electronic device

This case is about the establishment of an image processing model and an image processing method, and in particular, about a method for establishing an image defect detection model, a method for detecting defective images, and an electronic device for executing these methods.

Image recognition technology is widely used in defect detection, which improves the detection speed and accuracy compared to traditional manual detection. Defect detection technology relies on machine vision to compare the image to be read and the known defect image, so the defect features in the defect image must be defined in advance. Based on different application scenarios, the content of the defect image will also vary with the image of the application field. Users need to define the corresponding defect features according to the application scenario, however, this approach takes a lot of time to prepare for the pre-operation.

In order to efficiently find defects in images, neural network algorithms are used to learn the defect features in images to achieve automatic image defect detection. Neural network algorithms can infer image features from a large number of unknown images input by themselves. However, there is no efficient method to infer defect features from unknown images during the training stage.

In addition, the existing auto-encoder can be used for image reconstruction, filtering out defects during the reconstruction process to produce a defect-free reconstructed image. However, if the number of samples in the training phase is insufficient or the training image does not have defects, the auto-encoder model cannot learn the defect characteristics. In other words, the defects in the test image input to the auto-encoder will still appear in the reconstructed image, resulting in incorrect output results of the auto-encoder, that is, the accuracy of the judgment results in the test phase is low.

Accordingly, in the method of using a neural network algorithm to detect defects in an image, the problem of poor accuracy in judging defects in the image remains to be solved.

According to an embodiment of the present invention, a method for establishing an image defect detection model is disclosed. The image defect detection model is used to detect whether an image is abnormal. The method for establishing the image defect detection model includes: a) using a flawless sample image and a flawed sample image to generate a first training image; b) inputting the first training image to an automatic encoder to output a reconstructed image and calculate a first reconstruction error; c) calculating the difference between the first training image and the reconstructed image to generate a reconstructed error image; d) using the first training image, the reconstructed image, and the reconstructed error image to generate a second training image; e) inputting the first training image to an automatic encoder to output a reconstructed image and calculate a first reconstruction error; The second training image is fed to an image segmentation network to generate a mask image and calculate a second reconstruction error; f) a total reconstruction error is calculated based on the first reconstruction error and the second reconstruction error; g) the parameters of the automatic encoder and the image segmentation network are optimized based on the total reconstruction error; and h) steps a) to g) are repeatedly performed to complete the establishment and parameter optimization of the image defect detection model, wherein the image defect detection model includes the automatic encoder and the image segmentation network.

According to an embodiment of the present invention, a method for detecting defective images is disclosed, including: a) obtaining an image to be detected; b) inputting the image to be detected into an automatic encoder and outputting a reconstructed image; c) calculating the difference between the image to be detected and the reconstructed image to generate a reconstructed error image; d) generating a test image according to the image to be detected, the reconstructed image and the reconstructed error image; e) inputting the test image into an image segmentation network to generate a defect mask test image; and f) judging whether the image to be detected is a defective image according to an indicated area size of the defect mask test image.

The image processing method proposed in this case can automatically expand the images used for training, and in the subsequent testing phase of whether there are defects in the image, it can greatly improve the accuracy of judging whether there are defects in the image and the location of the defects in the image.

10: Electronic devices

110: Storage media

120: Processor

150: Image defect detection model

155: Automatic encoder

157: Image segmentation network

159: Computation module

302: flawless sample image

304: Defective sample image

310: First training video

312: Reconstructing the image

314: Defects

316: Reconstructing erroneous images

318: Defects

320: Second training video

328: Defects

334: Defect mask image

336: Filter area

338: Reserved area

E1: First error function

E2: Second error function

ES: Weighted formula

K1: First reconstruction error

K2: Second reconstruction error

KS: Total reconstruction error

S210~S280, S1110~S1180: Steps

FIG1 is a block diagram of an electronic device according to an embodiment of the present invention.

FIG2 is a flow chart of a method for establishing an image defect detection model according to an embodiment of the present invention.

Figure 3 is a schematic diagram of generating a first training image according to an embodiment of the present invention.

FIG4 is a schematic diagram of an output reconstructed image drawn according to an embodiment of the present invention.

FIG5 is a schematic diagram of calculating the first reconstruction error according to an embodiment of the present invention.

FIG6 is a schematic diagram of a reconstructed erroneous image generated according to an embodiment of the present invention.

FIG7 is a schematic diagram of generating a second training image according to an embodiment of the present invention.

FIG8 is a schematic diagram of generating a defect mask image according to an embodiment of the present invention.

FIG9 is a schematic diagram of calculating the second reconstruction error according to an embodiment of the present invention.

FIG10 is a schematic diagram showing the calculation of the total reconstruction error according to an embodiment of the present invention.

FIG11 is a flow chart of a defect image detection method according to an embodiment of the present invention.

The following further describes the present invention in combination with the drawings and embodiments, so that relevant personnel in the technical field to which the present invention belongs can better understand the present invention and implement it accordingly, but the embodiments are not intended to limit the present invention.

As used herein, terms such as "first" and "second" describe various elements, components, regions, layers and/or parts, which should not be limited by these terms. These terms may only be used to distinguish one element, component, region, layer or part from another. Unless the context clearly indicates otherwise, the terms such as "first" and "second" used herein do not imply a sequence or order.

Please refer to FIG. 1 , which is a block diagram of an electronic device according to an embodiment of the present invention. The electronic device 10 includes a storage medium 110 and a processor 120. The storage medium 110 is coupled to the processor 120. The electronic device 10 is used to establish an image defect detection model 150.

In one embodiment, the image defect detection model 150 is used to detect whether the image is abnormal.

The storage medium 110 is configured to store the image defect detection model 150. In one embodiment, the image defect detection model 150 is implemented using a neural network model architecture. The image defect detection model 150 includes an auto encoder 155, an image segmentation network 157, and a computing module 159.

In one embodiment, the autoencoder 155 can be a convolutional autoencoder (CNN AutoEncoder), a sparse autoencoder (Sparse AutoEncoder), a denoising autoencoder (Denoising AutoEncoder) or other autoencoders implemented in a multi-layer neural network architecture using an unsupervised learning algorithm.

In one embodiment, the image segmentation network 157 can be an object detection algorithm (such as U-Net or Mask RCNN algorithm) for detecting objects in an image and marking the object outline in the image and generating a mask image.

The computing module 159 is used to execute other neural network operations other than the automatic encoder 155 and the image segmentation network 157.

In one embodiment, the image defect detection model 150 includes an automatic encoder 155, an image segmentation network 157, and an operation module 159. In one embodiment, the automatic encoder 155, the image segmentation network 157, and the operation module 159 are software models, which are implemented by multiple program codes, so that the processor 120 loads the multiple program codes and executes the operations of the automatic encoder 155, the image segmentation network 157, and the operation module 159.

In one embodiment, the image defect detection model 150 can be a deep learning algorithm, such as a convolutional neural network (CNN), a recurrent neural network (RNN), a generative adversarial network (GAN), a multilayer perceptron (MLP), a deep Boltzmann machine (DBM), or a long short-term memory network (LSTM).

In one embodiment, the processor may be, but is not limited to, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a central processing unit (CPU), a system on chip (SoC), a field programmable gate array (FPGA), a network processor chip, or a combination of the above components.

In one embodiment, the storage medium may be, but is not limited to, a random access memory (RAM), a flash memory (Flash memory), a read-only memory (ROM), a hard disk drive (HDD), a solid state drive (SSD), an optical storage device, or a combination of the above elements.

In one embodiment, the electronic device 10 may be, but is not limited to, a tablet computer, a laptop computer, a personal computer, a desktop computer, a mainframe computer system, a computer workstation, a video computer, or other electronic devices equipped with a processor and/or a storage device that can execute applications installed thereon and/or in the cloud or can execute instructions stored locally or in the cloud.

In one embodiment, the image defect detection model 150 can be applied to, for example, factory production line product inspection, wafer inspection, printed circuit board inspection, public field anomaly inspection, or any field that uses machine vision to achieve purposeful inspection.

Please refer to FIG. 2, which is a flow chart of a method for establishing an image defect detection model according to an embodiment of the present invention. The method for establishing an image defect detection model can be executed by the electronic device 10 of FIG. 1. The processor 120 of the electronic device 10 loads multiple program codes in the storage medium 110 to execute multiple operations to achieve the establishment and training of the image defect detection model 150.

In one embodiment, the user can obtain multiple flawless sample images in advance for the purpose of detection. The flawless sample image refers to an image whose content is normal or correct based on the field or application. For example, in the application of printed circuit boards, the flawless sample image refers to an image of a printed circuit board with correct circuit components, correct positions of circuit components, and correct wiring. In another embodiment, the flawless sample image can be an image used for training and learning.

The flawless sample image can be stored in advance in an image collection of the storage medium 110 (not shown).

In step S210, the computing module 159 uses the flawless sample image and the flawed sample image to generate the first training image.

In one embodiment, the first training image is generated by synthesizing a flawless sample image and a flawed sample image.

Please refer to Figure 3, which is a schematic diagram of generating the first training image according to an embodiment of the present invention.

As shown in FIG3 , the flawless sample image 302 and the flawed sample image 304 are synthesized to generate a first training image 310. The first training image 310 includes image features that are the same or similar to those of the flawless sample image 302 and the flawed sample image 304.

In one embodiment, the flaw sample image 304 includes a noise sample image and a defect sample image.

In one embodiment, the computing module 159 generates image noise using a random function, and performs affine transformation, perspective transformation, color transformation or a combination thereof to process the image noise and then synthesizes it with a flawless sample image to generate a noise sample image.

In one embodiment, the computing module 159 performs affine transformation, perspective transformation, color transformation or a combination thereof to process any image that can be regarded as a defect and then synthesizes it with a flawless sample image to generate a defect sample image.

In one embodiment, the image set stored in the storage medium 110 includes a plurality of flawless sample images and a plurality of defective sample images. There is a ratio between the number of defective sample images and the number of flawless sample images. For example, the ratio is 1 to 100, that is, the computing module 159 takes the same flawed sample image and 100 flawless sample images in the image set to generate 100 first training images respectively. This ratio is used to adjust the difference between the first training image and the flawless sample image.

In step S220, the computing module 159 inputs the first training image to the automatic encoder 155, and the automatic encoder 155 outputs the reconstructed image and calculates the first reconstruction error.

Please refer to Figure 4, which is a schematic diagram of the output reconstructed image drawn according to an embodiment of the present case.

As shown in FIG. 4 , the input image of the automatic encoder 155 is the first training image 310. The encoder (not shown) and decoder (not shown) of the automatic encoder 155 perform image processing on the first training image 310 and output a reconstructed image 312 of the flawless sample image 302.

In one embodiment, the automatic encoder 155 has the ability to process defects in the first training image 310 (e.g., filter defects 314), and convert the first training image 310 into an image that is the same as or similar to the defect-free sample image 302, that is, the reconstructed image 312. The reconstructed image 312 is the output image of the automatic encoder 155.

According to the above embodiment, in step S220, the electronic device 10 further calculates the first reconstruction error. In one embodiment, the electronic device 10 inputs the pixel value of the flawless sample image 302 and the pixel value of the reconstructed image 312 into the first error function to calculate the first reconstruction error.

，其中X _i 為無瑕疵樣本影像302的像素值，Y _i 為重構影像312的像素值。 In one embodiment, the first error function is:

, where Xi is the pixel value of the flawless sample image 302, _and Yi _is the pixel value of the reconstructed image 312.

Please refer to Figure 5, which is a schematic diagram of calculating the first reconstruction error according to an embodiment of the present invention.

The electronic device 10 calculates the square sum of the difference between the pixel value of the flawless sample image 302 and the pixel value of the reconstructed image 312 for each pixel using the first error function E1, and then calculates the average of the square sum to obtain the first reconstruction error K1. The first reconstruction error K1 presents the overall error value or difference between the flawless sample image 302 and the reconstructed image 312, reflecting the degree of noise processing (elimination) by the automatic encoder 155 at present.

The first reconstruction error K1 can be fed back to the automatic encoder 155 as one of the factors for adjusting the parameters of the automatic encoder 155 to optimize the ability of the automatic encoder 155 to process noise.

In step S230, the computing module 159 calculates the difference between the first training image and the reconstructed image to generate a reconstructed error image.

Please refer to Figure 6, which is a schematic diagram of the reconstructed erroneous image generated according to an embodiment of the present invention.

As shown in FIG6 , the difference between the reconstructed image 312 and the first training image 310 is the reconstructed error image 316.

In one embodiment, the first training image 310 includes a defect 314, and the reconstructed error image 316 includes a defect 318. The defect 314 is a defect feature before image processing, and the defect 318 is a defect feature after image processing. The defect 318 is similar to the defect 314, so the reconstructed error image 316 can be regarded as another defect image similar to the defect sample image 304 (as shown in FIG. 3).

In step S240, the computing module 159 generates a second training image using the first training image 310, the reconstructed image 312, and the reconstructed error image 316.

Please refer to Figure 7, which is a schematic diagram of generating a second training image according to an embodiment of the present invention.

As shown in FIG. 7 , the computing module 159 combines or overlaps the first training image 310 , the reconstructed image 312 , and the reconstructed error image 316 to generate a multi-dimensional image (i.e., the second training image 320 ).

The second training image 320 is generated by merging or superimposing the first training image 310, the reconstructed image 312, and the reconstructed error image 316. Therefore, the image features of the second training image 320 include a defect 328, which is related to the defect 314 in the first training image 310 and the defect 318 in the reconstructed error image 316.

In step S250, the computing module 159 inputs the second training image to the image segmentation network 157, the image segmentation network 157 generates a mask image and the computing module 159 calculates the second reconstruction error.

Please refer to Figure 8, which is a schematic diagram of generating a defect mask image according to an embodiment of the present invention.

As shown in FIG8 , the second training image 320 is input to the image segmentation network 157. The image segmentation network 157 classifies the pixels of the second training image 320 as defects or background according to the pixels and image features of the second training image 320.

In one embodiment, the image segmentation network 157 generates a defect mask image 334 after classifying all pixels of the second training image 320. The defect mask image 334 includes a reserved area 338 and a filter area 336. The reserved area 338 corresponds to the defect 328 of the second training image 320. The filter area 336 corresponds to the background of the second training image 320.

In one embodiment, the computing module 159 uses the defect mask image 334 to perform image processing, filters out the image block corresponding to the filter area 336 of the processed image, and leaves the image block corresponding to the reserved area 338. In this embodiment, the left image block is used to indicate the defect block of the processed image.

In step S250, the calculation module 159 further calculates the second reconstruction error. In one embodiment, the calculation module 159 inputs the defect sample image 304 and the defect mask image 334 to the second error function E2 to calculate the second reconstruction error K2.

In one embodiment, the second error function is: E2 = _-αt (1 _- pt ) ^γlog ( pt ), where pt is the _probability of correctly judging the processed image as a defect image or a non -defect image using the defect mask image ₃₃₄ , _αt and γ are coefficients of the second error function E2, which are hyperparameters used for subsequent feedback adjustment of the weight values of the image segmentation network 157.

Please refer to Figure 9, which is a schematic diagram of calculating the second reconstruction error according to an embodiment of the present case.

The operation module 159 calculates the difference between the defect sample image 304 and the defect mask image 334 using the second error function E2 to obtain the second reconstruction error K2. The second reconstruction error K2 presents the overall error value or difference between the defect sample image 304 and the defect mask image 334.

In step S260, the total reconstruction error is calculated based on the first reconstruction error K1 and the second reconstruction error K2.

In one embodiment, the computing module 159 calculates a weighted formula to obtain a total reconstruction error.

In one embodiment, the weighting formula is: ES = w ₁ E 1 + w ₂ E 2, where w ₁ and w ₂ are weighting coefficients (eg, a ratio whose total is 100%).

Please refer to Figure 10, which is a schematic diagram of calculating the total reconstruction error according to an embodiment of the present invention.

As described above, the first reconstruction error K1 calculated in step S220 and the second reconstruction error K2 calculated in step S250 are respectively input into the weighting formula ES. The operation module 159 uses the weighting formula ES to proportionally distribute the first reconstruction error K1 and the second reconstruction error K2 with weighting coefficients w_1 and w_2, respectively, and calculates the total reconstruction error KS after adding them up.

In step S270, the computing module 159 optimizes the parameters of the automatic encoder 155 and the image segmentation network 157 according to the total reconstruction error.

In one embodiment, the total reconstruction error KS can be fed back to the auto-encoder 155 and the image segmentation network 157 for parameter optimization via back propagation.

After the parameters of the automatic encoder 155 and the image segmentation network 157 are adjusted, the electronic device 10 executes the aforementioned steps S210 to S270 again, continuously training the automatic encoder 155 and the image segmentation network 157 and adjusting the parameters of the automatic encoder 155 and the image segmentation network 157 to improve the accuracy of the automatic encoder 155 and the image segmentation network 157.

In step S280, the electronic device 10 completes the image training and parameter optimization of the automatic encoder 155 and the image segmentation network 157.

In one embodiment, the electronic device 10 determines whether the total reconstruction error KS is less than a threshold value each time. If the total reconstruction error KS is less than the threshold value, it is determined that the establishment and parameter optimization of the image defect detection model 150 have been completed.

After establishing and optimizing the image defect detection model 150, the image defect detection model 150 can be applied to detect whether the image to be detected is a defective image.

Please refer to FIG. 11 , which is a flow chart of a defect image detection method according to an embodiment of the present invention. The defect image detection method can be executed by the electronic device 10 of FIG. 1 . The processor 120 of the electronic device 10 loads multiple program codes in the storage medium 110 to execute multiple operations to realize the detection of defect images.

In step S1110, the electronic device 10 obtains the image to be detected.

In one embodiment, the image to be inspected can be a factory production line product image, a wafer image, a printed circuit board image, a public field image, or an image taken from any field.

In one embodiment, the electronic device 10 includes an image capture module (such as a camera) for photographing the object to be tested to obtain an image to be tested.

In step S1120, the computing module 159 inputs the image to be detected to the automatic encoder 155, and the automatic encoder 155 outputs the detected reconstructed image.

In one embodiment, the automatic encoder 155 performs encoder operations and decoder operations on the image to be detected and outputs a detected reconstructed image of the image to be detected.

In this embodiment, the automatic encoder 155 has been optimized, so it can filter out the defects of the image to be detected in this step. In other words, the detected reconstructed image is a more accurate image than the image to be detected.

In step S1130, the computing module 159 calculates the difference between the image to be detected and the detected reconstructed image to generate a detected reconstructed error image.

The detection and reconstruction error image is the difference image between the image to be detected and the detection and reconstruction image. In other words, the purpose of this step is to extract the image block that is determined to be a defect of the image to be detected and represent it with the detection and reconstruction error image.

In step S1140, the computing module 159 generates a test image based on the image to be detected, the detected reconstructed image, and the detected reconstructed error image.

In one embodiment, the computing module 159 combines or overlaps the image to be detected, the detected reconstructed image, and the detected reconstructed error image to generate a multi-dimensional image (i.e., a test image).

Since there are objective defect blocks in the image to be detected, the detected reconstructed image is an image with defects filtered out (an image without defects considered by the automatic encoder 155), and the detected reconstructed error image is an image belonging to the defect block of the image to be detected (an image considered by the automatic encoder 155), the test image generated by merging or superimposing these three images can highlight the intersection of the objective defects and the defects determined by the automatic encoder 155.

In step S1150, the computing module 159 inputs the test image to the image segmentation network 157, and the image segmentation network 157 generates a defect mask test image.

In one embodiment, the test image is input to the image segmentation network 157. The image segmentation network 157 classifies multiple pixel blocks of the test image into defect areas or background areas based on the pixels and image features of the test image.

In one embodiment, the image segmentation network 157 generates a defect mask test image after classifying all pixel blocks of the test image. The defect mask test image includes a defect area and a background area. The defect area corresponds to the image block of the defect of the test image. The background area corresponds to the image block of the non-defect (background) of the test image.

In step S1160, the computing module 159 determines whether the indicated area of the defect mask test image is greater than the valve value. If the indicated area is equal to or less than the valve value, step S1170 is executed. If the indicated area is greater than the valve value, step S1180 is executed.

In one embodiment, the indicated area is the area of the defect region of the defect mask test image.

In step S1170, since the indicated area is less than or equal to the valve value, it means that the defect amount of the image to be tested does not exceed the tolerance value, so the calculation module 159 determines that the image to be tested is a flawless image.

In step S1180, since the indicated area is larger than the valve value, it means that the defect amount of the image to be tested exceeds the tolerance value, so the calculation module 159 determines that the image to be tested is a defective image.

In one embodiment, the electronic device 10 will send a notification to the user (e.g., a notification that the appearance of an item is abnormal) so that the user can perform corresponding processing.

In summary, this case proposes to establish an image defect detection model and use the image defect detection model to detect defective images, and improve the diversity of the training model by adding noise samples and bad samples to the detection model. In the detection stage, the detection results include the position and size of the defect in the image, and a notification is issued based on the detection results to determine whether the image to be detected has defects (i.e., whether the object to be detected is abnormal). The model's parameter optimization is improved through image processing operations composed of multiple steps. The image defect detection model established according to the method of this case can improve the accuracy of detecting defective images.

The above is only a specific example of this case, and does not limit the scope of the patent application of this case. Therefore, all equivalent changes made by applying the content of this case are also included in the scope of this case and should be stated.

S210~S280: Steps

Claims (15)

A method for establishing an image defect detection model, wherein the image defect detection model is used to detect whether an image is abnormal, the method comprising: a) using a flawless sample image and a flawed sample image to generate a first training image; b) inputting the first training image to an automatic encoder to output a reconstructed image and calculating a first reconstruction error; c) calculating the difference between the first training image and the reconstructed image to generate a reconstruction error image; d) using the first training image, the reconstructed image and the reconstruction error image to generate a second training image; e) inputting the second training image into an image segmentation network to generate a mask image and calculate a second reconstruction error; f) calculating a total reconstruction error according to the first reconstruction error and the second reconstruction error; g) optimizing the parameters of the automatic encoder and the image segmentation network according to the total reconstruction error; and h) repeatedly executing steps a) to g) to complete the establishment and parameter optimization of the image defect detection model, wherein the image defect detection model includes the automatic encoder and the image segmentation network. As described in claim 1, before step a), the method includes: using a random function to generate a noise sample image and synthesizing the noise sample image and the flawless sample image to obtain the flawed sample image; or transforming a bad sample image and synthesizing the transformed bad sample image and the flawless sample image to obtain the flawed sample image. As described in claim 1, step b) includes: Inputting the pixel value of the flawless sample image and the pixel value of the reconstructed image into a first error function for calculation to obtain the first reconstruction error. As described in claim 1, step d) includes: merging the first training image, the reconstructed image and the reconstructed error image to generate the second training image. The method as claimed in claim 1, wherein step e) comprises: classifying the pixels of the second training image through the image segmentation network to obtain a reserved area corresponding to the defect and a filtered area corresponding to the background; and outputting the mask image including the reserved area and the filtered area. As described in claim 5, step e) includes: inputting the pixel value of the defect sample image and the pixel value of the mask image into a second error function for calculation to obtain the second reconstruction error. As described in claim 6, step f) includes: using a first weight and a second weight to perform weighted calculation on the first reconstruction error and the second reconstruction error respectively to obtain the total reconstruction error. The method as claimed in claim 7, wherein step g) comprises: adjusting the parameters of the automatic encoder and the image segmentation network according to the parameter value of the second error function, the first weight and the second weight. As described in claim 8, step h) includes: repeatedly executing steps a) to g) until the total reconstruction error is less than a threshold value to complete the establishment of the image defect detection model and parameter optimization. An electronic device for establishing an image defect detection model comprises: a storage medium configured to store the image defect detection model; and a processor coupled to the storage medium and configured to execute a method as recited in any one of claims 1 to 9. A method for detecting defective images using the image defect detection model established by the method of any one of claims 1 to 9, comprising: a1) obtaining an image to be detected; b1) inputting the image to be detected into the automatic encoder and outputting a detection reconstructed image; c1) calculating the difference between the image to be detected and the detection reconstructed image to generate a detection reconstructed error image; d1) generating a test image based on the image to be detected, the detection reconstructed image and the detection reconstructed error image; e1) inputting the test image into the image segmentation network to generate a defect mask test image; and f1) judging whether the image to be detected is a defective image based on an indicated area size of the defect mask test image. As described in claim 11, step d1) includes: merging the image to be detected, the detected reconstructed image and the detected reconstructed error image to generate the test image. The method as claimed in claim 11, wherein step e1) comprises: classifying the pixels of the test image through the image segmentation network to obtain a defect area and a background area; and outputting the defect mask test image including the defect area and the background area. The method of claim 12, wherein step f1) includes: calculating the area of the defect region of the defect mask test image to obtain the indication area; if the indication area is greater than a valve value, judging that the image to be tested is the defect image; and if the indication area is less than or equal to the valve value, judging that the image to be tested is a non-defective image. An electronic device for detecting defective images, comprising: a storage medium configured to store the defective image detection model; and a processor coupled to the storage medium and configured to execute the method of claim 11.

Images