CN111340785A - Model training method, product surface defect detection method and storage medium - Google Patents

Abstract

The invention discloses a model training method, a product surface defect detection method and a storage medium. The training method comprises the steps of obtaining an automatic coding network, training the automatic coding network, and obtaining and generating a confrontation network; the generation countermeasure network comprises a generator network and a discriminator network, the discriminator network is constructed by an encoder network in the automatic encoding network, and the generation countermeasure network is trained. The convolutional automatic coding network-generation countermeasure network combined network obtained by training has the advantages of unsupervised training of the convolutional automatic coding network and semi-supervised training of the generation countermeasure network, has better generalization capability, can adapt to the conditions of few training samples, complex image appearance, large group difference and the like faced by the occasions of detecting the surface defects of products, particularly the surface defects of steel products, and obtains better recognition effect. The invention is widely applied to the technical field of image detection.

Description

Model training method, product surface defect detection method and storage medium

technical field

The invention relates to the technical field of image detection, in particular to a model training method, a product surface defect detection method and a storage medium.

Background technique

In the production and maintenance process of steel products and other products, the detection and analysis of surface defects is an efficient defect detection method. However, identifying surface defects has always been a difficult task due to the rare occurrence of defects and changes in appearance.

In recent years, deep learning methods have shown excellent performance in image classification, especially when there are enough training samples. Therefore, some existing technologies for identifying and detecting product surface defects have emerged one after another, including extreme learning machine (ELM) and support vector machine (SVM), etc. On this basis, genetic algorithm (GA), RNAMlet feature corrector, scale Invariant feature transformation (SIFT) and shear wave transformation, etc. are used for improvement. The general principle is to shoot the surface of the product, input the captured image into the trained deep learning model, and obtain the output of the deep learning model. Identify the results to determine whether there are defects on the surface of the product or what type of defects exist. Therefore, the effectiveness of this approach depends on the performance of the deep learning model.

However, product surface defects have a complex appearance, which makes the resulting pictures also highly complex. Figure 1 shows the surface defects of iron and steel products, in which parts a-d are seams, and parts e-h are oxide scales. It can be seen from Figure 1 that even the same surface defects have great differences in appearance, that is, the surface of the product is very different. Defective images are characterized by large intra-group differences, and the image background is very complex, which makes the generalization ability of the existing technology poor. When applying the existing technology, the model can only be constructed and trained separately for each appearance. Greatly increases the cost of use and reduces the efficiency of use. Another feature of product surface defect images is that it is difficult to obtain a sufficient number of sample images to train the model, and the context of product surface defect images is very different from that of most pre-trained models, which makes transfer learning suitable for small dataset scenarios. The improved technology is difficult to apply in the field of product surface defect detection.

SUMMARY OF THE INVENTION

In view of at least one of the above technical problems, the purpose of the present invention is to provide a model training method, a product surface defect detection method and a storage medium.

On the one hand, an embodiment of the present invention includes a training method for a product surface defect detection model, comprising the following steps:

Obtain an auto-encoding network; the auto-encoding network includes an encoder network and a decoder network;

obtaining a sample image; at least a portion of the sample image includes product surface defects;

training the auto-encoding network using the sample images;

Obtain a generative adversarial network; the generative adversarial network includes a generator network and a discriminator network, and the discriminator network is constructed by the encoder network in the auto-encoding network;

Obtain a real image and its mark; the real image includes product surface defects, and the mark is used to indicate the type of product surface defect included in the corresponding real image;

Input the real image or the fake image generated by the generator network into the discriminator network, obtain the output result of the discriminator network, adjust the parameters of the discriminator network and/or the generator network, until The loss function of the discriminator network and/or the loss function of the generator network reaches the target value;

The trained generative adversarial network is used as a product surface defect detection model.

Further, the step of adjusting the parameters of the discriminator network and/or the generator network specifically includes:

reconstructing sample images by the decoder network;

Obtain the error of the decoder network in the reconstruction process;

The encoder network is adjusted according to the error.

Further, the loss function of the discriminator network is:

Further, the discriminator network is obtained by connecting the encoder network and the multi-classification network.

Further, the multi-classification network is a softmax network.

Further, the automatic coding network is a convolutional automatic coding network; the encoder network includes a plurality of convolutional layers and a passthrough layer; the passthrough layer is used to size the feature map output by the shallower convolutional layer. downscale and concatenate with feature maps output by deeper convolutional layers.

Further, the deeper convolutional layer is the last convolutional layer in the encoder network.

Further, at least one of the convolutional layers in the encoder network performs a max-pooling operation.

On the other hand, an embodiment of the present invention also includes a method for detecting surface defects of a product, comprising the following steps:

Obtain an image to be detected; the image to be detected includes the product surface;

Input the image to be detected into the product surface defect detection model; the product surface defect detection model is trained by the training method described in the embodiment;

The output result of the product surface defect detection model is acquired, and the product surface defect type is determined according to the output result.

On the other hand, an embodiment of the present invention further includes a storage medium, in which processor-executable instructions are stored, and when executed by the processor, the processor-executable instructions are used to perform the object shape measurement described in the embodiments method.

The beneficial effects of the present invention are: the generative adversarial network trained by the model training method in the embodiment is actually a convolutional auto-encoding network-generative adversarial network joint network, which has both the unsupervised convolutional auto-encoding network. The advantages of training, as well as the advantages of semi-supervised training of generative adversarial networks, have good generalization ability, and can adapt to the surface defects of products, especially the surface defects of steel products. There are few training samples, complex image appearance, and intra-group differences. In large and other situations, a better recognition effect is achieved.

Description of drawings

Figure 1 is an image of surface defects of a steel product;

Fig. 2 is the principle schematic diagram of the model training method in the embodiment;

Fig. 3 is the principle schematic diagram of the convolutional auto-encoding network constructed in the embodiment;

4 is a schematic structural diagram of a convolutional auto-encoding network constructed in an embodiment;

5 is a schematic diagram of the principle of upsampling performed in a decoder network in an embodiment;

FIG. 6 is a schematic diagram of the principle of a generative adversarial network in an embodiment.

Detailed ways

Example 1

A training method proposed in this embodiment includes the following steps:

P1. Obtain a convolutional auto-encoding network; the convolutional auto-encoding network includes an encoder network and a decoder network;

P2. Obtain a sample image; at least part of the sample image includes product surface defects;

P3. Use the sample images to train the convolutional auto-encoding network;

P4. Obtain a generative adversarial network; the generative adversarial network includes a generator network and a discriminator network, and the discriminator network is constructed by the encoder network in the convolutional auto-encoding network;

P5. Obtain a real image and its mark; the real image includes product surface defects, and the mark is used to indicate the type of product surface defect included in the corresponding real image;

P6. Input the real image or the fake image generated by the generator network to the discriminator network, obtain the output result of the discriminator network, and adjust the parameters of the discriminator network and/or generator network , until the loss function of the discriminator network reaches the target value; the loss function is determined by the output result of the discriminator network and the label of the real image or the label of the fake image;

The trained generative adversarial network is used as a product surface defect detection model.

On the basis of not affecting the logic, the sequence of steps P1-P6 can be adjusted, which does not affect the implementation of the training method in this embodiment.

The goal of steps P1-P6 is to build a generative adversarial network and train the generative adversarial network. The trained generative adversarial network is the product surface defect detection model we want to get.

The principle of steps P1-P6 is shown in FIG. 2 . Preferably, the auto-encoding network used in this embodiment is a convolutional auto-encoding network (CAE). Compared with the standard auto-encoding network, the convolutional auto-encoding network is more suitable for high-dimensional image data. Also, in convolutional autoencoder networks, weights and biases are shared across all positions in the input, which means that convolutional autoencoders can preserve spatial locality. Therefore, the reconstruction of the decoder network is based on a linear combination of the basic image patches output by the encoder network.

By executing steps P1-P4, the convolutional auto-encoding network is first constructed, and the convolutional auto-encoding network is trained using sample images. For the trained convolutional auto-encoding network, the encoder network part is directly or slightly modified. , used as the discriminator network in the generative adversarial network, together with the constructed generator network, constitutes the generative adversarial network.

In steps P1-P3, the principle of the constructed convolutional auto-encoding network is shown in Figure 3, where the input data x represents an m-dimensional vector, where x ∈ Rm. The output data features represent n-dimensional vector features, where n∈Rn and m>n. The principle of convolutional auto-encoding networks is to first transform the input data into a usually lower dimensional space in the encoder network, and then expand to reproduce the original data in the decoder network.

The convolutional autoencoder network runtime consists of three main steps:

Step 1: Convert the input x to the code of the encoder network by:

f=sigmoid(W ₁ ^T x+b ₁ );

where x is the input vector; W ₁ ^T is the weight matrix between the input and hidden layers; b ₁ is the bias vector; and f is the output value of the encoder network.

Step 2: Based on the output of the encoder network, the decoder network reconstructs the input value by:

是隐藏层和输出层之间的权重矩阵；b₂偏差矢量。

是解码器网络的输出值。where f is the output of the encoder network;

is the weight matrix between the hidden and output layers _; b is the bias vector.

is the output value of the decoder network.

The training of the convolutional auto-encoding network is unsupervised, that is, there is no need to mark the sample images obtained in step P2, only a part of the sample images need to include product surface defects, and the other part of the sample images do not include product surface defects. The goal of convolutional autoencoder network training is to minimize the error between the original image and the output of the decoder network, which is:

是解码器网络的输出，

通常，还可以在

后加惩罚项α以强制卷积自动编码网络对输入图像的稀疏表示进行编码。本实施例中，惩罚项α满足：where En represents the encoder network; De represents the decoder network; x is the input vector;

is the output of the decoder network,

Usually, you can also

A penalty term α is added to force the convolutional auto-encoding network to encode a sparse representation of the input image. In this embodiment, the penalty term α satisfies:

是编码器网络的平均输出；ρ是实验中设定的目标值0.05；KL是Kullback-Leibler散度，它有助于编码器网络输出更多的稀疏值。

is the average output of the encoder network; ρ is the target value 0.05 set in the experiment; KL is the Kullback-Leibler divergence, which helps the encoder network to output more sparse values.

The more detailed structure of the convolutional auto-encoding network used is shown in Figure 4, where the specific parameters of the encoder network are shown in Table 1, and the specific parameters of the decoder network are shown in Table 2.

Table 1

Table 2

The encoder network includes multiple convolutional layers, and at the end of each convolutional layer, it also includes a max-pooling layer to perform a max-pooling operation on the data output by the convolutional layer.

In this example, the encoder network takes an image of size 224×224 as input and is processed by eight convolutional layers and four max-pooling layers to construct a set of discriminative feature maps. A convolutional layer in an encoder network can be described as:

y=f(∑f _c *k _c +b _c );

where _y represents the output feature map of the convolutional layer; fc and _kc are the cth feature map and the cth convolution kernel, respectively. b _c is the c-th deviation; * denotes a convolution operation. The training of the convolutional autoencoder network is to continuously learn k _c and b _c .

In this embodiment, a rectified linear unit (ReLU) is used in each convolutional layer of the encoder network, which can improve the nonlinear expression ability of the convolutional auto-encoding network. A rectified linear unit is represented as:

y=max(x,0);

where x is the output of the convolutional layer; the max pooling layer can reduce the dimension of the feature map, and a max pooling layer can be expressed as:

where x(i,j) represents the (i,j)th merged region; R is all the different merged regions; y(i,j) is the output of the (i,j)th merged region; the maximum merge increases Translation-invariant convolutional auto-encoding networks and making them encode sparser features.

Since the size of the input image to the convolutional auto-encoding network is 224×224, and there are four max-pooling layers in the encoder network, the output feature map of the encoder network is of size 14×14, which is then substituted into the decoding The network is used to reconstruct the original image picture as the final output of the convolutional auto-encoding network.

In this embodiment, a passthrough layer is set in the encoder network. As shown in Figure 4, the input of the passthrough layer is connected to the output of a shallower convolutional layer in the encoder network, where the depth is a description of the position of the convolutional layer in the encoder network, and the deeper volume The convolutional layers follow the shallower convolutional layers. The passthrough layer obtains the feature map output by this shallower convolutional layer, and stacks adjacent features in this feature map into different feature map channels, so that this feature map is converted into multiple feature maps of smaller size For example, a passthrough layer can convert a feature map of size 4×4 output by a shallower convolutional layer into 4 different feature maps of size 2×2 each.

The output of the passthrough layer is connected to a deeper convolutional layer in the encoder network. After the above processing is completed, the passthrough layer combines these feature maps output by the shallow convolutional layer with those outputted by the deeper convolutional layer in the depth channel. The feature maps are concatenated to form the final output of the encoder network.

In this embodiment, the shallower convolutional layer connected to the passthrough layer is the fifth convolutional layer in the encoder network. This convolutional layer outputs 128 different feature maps with a size of 28×28. Passthrough The layer converts these 128 feature maps into 512 feature maps each of size 14×14. The shallower convolutional layer connected by the passthrough layer is the last convolutional layer in the encoder network, which itself can output 128 feature maps with a size of 14×14. The passthrough layer will convert the obtained feature map with the last volume. The feature maps output by the multi-layer layer are connected to obtain 640 feature maps with a size of 14 × 14, that is, the final output of the encoder network is 14 × 14 × 640.

The passthrough layer can retain the detailed features in the shallower convolutional layers, that is, it helps the encoder network to better extract the fine-grained features in the input data, so that the final output of the convolutional auto-encoding network still contains a large number of detailed features, avoiding Increase the resolution of the encoder network due to loss of details due to too many convolutional layers.

The decoder network takes the output of the encoder network as input and is convolved with nine convolutional and upsampling layers. The principle of upsampling performed in the decoder network is shown in Figure 5. Upsampling the feature map by duplication, the process and effect of which are opposite to max pooling, can expand the size of the feature map.

After building and training the convolutional autoencoder network, the encoder network in the convolutional autoencoder network can be directly used as the discriminator network in the generative adversarial network.

Build a generator network. In this example, the structure of the generator network is shown in Table 3. It consists of 9 convolutional layers and 4 upsampling layers. The input size is 14×14, and the uniform value from 0 to 1 Distribution sampling. The role of the upsampling layer is the same as the role of the decoder network in the convolutional auto-encoding network, which is to expand the size of the feature map. After these upsampling and convolutions, the random noise input received by the generator network is mapped into a fake image.

table 3

The fake image is relative to the real image, that is, the image that does not belong to the real image, the fake image and the real image are input into the discriminator network, and the discriminator network outputs the judgment result of the fake image and the real image. Referring to Fig. 6, in the case where the encoder network in the convolutional auto-encoding network is directly used as the discriminator network, the judgment result of the discriminator network is binary, that is, the image input into the discriminator network is the product surface. The real image obtained by the real defect is still a fake image generated by random noise.

According to the judgment result of the discriminator, the specific value of the loss function of the discriminator network is determined; the loss function of the discriminator network can be defined as:

表示训练目标是使损失函数最大化。where D represents the discriminator network, G represents the generator network that takes random noise z as input, x represents the real images of the dataset,

Indicates that the training objective is to maximize the loss function.

The training goal of the generator network is to generate images with the largest possible value D(x) to test the discriminator network. The loss function of the generator can be defined as:

When the loss function does not reach the target value, the parameters of the generator network and the discriminator network are adjusted in the direction of gradient descent for the next round of processing. The two adjacent rounds of processing can adjust the parameters of the generator network and the discriminator network respectively to achieve the result of training in turn. When the loss function of the discriminator network achieves the target value, the training of the discriminator network can be ended, that is, The parameters of the discriminator network are no longer changed; when the loss function of the generator network achieves the target value, the training of the generator network can be ended, that is, the parameters of the generator network will not be changed. When the training of both the discriminator network and the generator network is completed, the training of the entire generative adversarial network is completed.

In this embodiment, the discriminator network in the generative adversarial network is obtained by connecting a multi-classification network on the basis of the encoder network of the convolutional auto-encoding network, and the multi-classification network can be selected as a softmax network. The multi-classification network can output multiple classification results, that is, in addition to identifying whether the received image is a real image or a fake image, it can also identify what kind of product surface defects the received image contains. Discriminator network with better generalization ability. At this time, in the corresponding training data used, the marking of the real image not only indicates that it belongs to the real image, but also indicates the type of product surface defect contained in it, so that the discriminator network can be trained to have the ability to identify the type of product surface defect. Ability.

When performing step P4 to use the encoder network as the discriminator network, the connection between the decoder network and the encoder network can be completely truncated. In this embodiment, the connection between the decoder network and the encoder network may also be reserved during the training of the generative adversarial network. At this time, the step of adjusting the parameters of the discriminator network and/or the generator network in step P4 consists of the following steps:

P401. Reconstruct a sample image by the decoder network;

P402. Obtain the error of the decoder network in the reconstruction process;

P403. Adjust the encoder network according to the error.

The loss function of the discriminator network accordingly becomes:

where De represents the decoder network; En represents the encoder network, which is also used as the convolutional layer of the discriminator network; α is used to reduce the loss weight of image reconstruction to ensure fast convergence of the discriminator network.

By preserving the connection between the decoder network and the encoder network, the decoder network still reconstructs sample images from the actual production line and propagates the error back to the encoder network, so that while training the generative adversarial network, the discriminator network still The features of the sample images can be learned from the decoder network to improve the recognition accuracy of the discriminator network.

In summary, steps P1-P6 actually include two stages. The first stage is the process of building and training the convolutional auto-encoding network described in steps P1-P3, and the second stage is the construction and training generation described in steps P4-P6. The process of adversarial networks. In actual execution, Keras or Tensorflow can be used to build convolutional auto-encoding networks and generative adversarial networks. The learning rates of the first and second stages are set to 0.001 and 0.0001, respectively, the momentum is set to 0.9, and the weight decay is set to 0.0005, a better recognition rate can be achieved.

The generative adversarial network constructed and trained by performing steps P1-P6 is actually a convolutional auto-encoding network-generative adversarial network joint network, which combines the advantages of unsupervised training of convolutional auto-encoding networks and the advantages of generative adversarial networks. The advantages of semi-supervised training are that it has good generalization ability, and can adapt to the situation of product surface defects, especially steel product surface defect detection, where there are few training samples, complex image appearance, and large differences within groups, etc., and achieve better results. Identify the effect.

Example 2

After performing the training method described in Embodiment 1, the resulting generative adversarial network is used as a product surface defect detection model, which is used to implement the following product surface defect detection methods:

S1. Obtain an image to be detected; the image to be detected includes the surface of the product, and the image to be detected can usually be obtained by shooting or scanning;

S2. Input the image to be detected into the product surface defect detection model;

S3. Obtain the output result of the product surface defect detection model, and determine the product surface defect type according to the output result.

Using the product surface defect detection model obtained in Example 1 to perform the product surface defect detection method, the same technical effect as described in Example 1 can be obtained, that is, it can be adapted to product surface defects, especially in steel product surface defect detection occasions Faced with the complex appearance of the image, large differences within the group, etc., a high recognition accuracy is achieved.

Example 3

The training method described in Embodiment 1 and the detection method described in Embodiment 2 are written into corresponding computer codes and written into the storage medium. When the storage medium is connected to the controller, the computer program codes therein can be read. Take it out and execute it, so that steps P1-P6 or S1-S3 are automatically executed to achieve the same technical effect as described in Embodiment 1 or Embodiment 2.

It should be noted that, unless otherwise specified, when a feature is called "fixed" or "connected" to another feature, it can be directly fixed or connected to another feature, or it can be indirectly fixed or connected to another feature. on a feature. In addition, descriptions such as upper, lower, left, right, etc. used in the present disclosure are only relative to the mutual positional relationship of each component of the present disclosure in the accompanying drawings. As used in this disclosure, the singular forms "a," "the," and "the" are intended to include the plural forms as well, unless the context clearly dictates otherwise. Also, unless otherwise defined, all technical and scientific terms used in this embodiment have the same meaning as commonly understood by those skilled in the art. The terms used in the description of the embodiments are only used to describe specific embodiments, rather than to limit the present invention. As used in this example, the term "and/or" includes any combination of one or more of the associated listed items.

It will be understood that, although the terms first, second, third, etc. may be used in this disclosure to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish elements of the same type from one another. For example, a first element could be termed a second element, and, similarly, a second element could be termed a first element, without departing from the scope of the present disclosure. The use of any and all examples or exemplary language ("for example," "such as," etc.) provided in this embodiment is intended only to better illustrate embodiments of the invention and does not detract from the scope of the invention unless otherwise requested impose restrictions.

It should be appreciated that embodiments of the present invention may be implemented or implemented by computer hardware, a combination of hardware and software, or by computer instructions stored in non-transitory computer readable memory. The methods can be implemented in a computer program using standard programming techniques - including a non-transitory computer-readable storage medium configured with a computer program, wherein the storage medium so configured causes the computer to operate in a specific and predefined manner - according to the specific Methods and figures described in the Examples. Each program may be implemented in a high-level process or target terminal-oriented programming language to communicate with a computer system. However, if desired, the program can be implemented in assembly or machine language. In any case, the language can be a compiled or interpreted language. Furthermore, the program can be run on a programmed application specific integrated circuit for this purpose.

Furthermore, the operations of the processes described in this embodiment may be performed in any suitable order unless otherwise indicated by this embodiment or otherwise clearly contradicted by context. The processes described in this embodiment (or variations and/or combinations thereof) may be performed under the control of one or more computer systems configured with executable instructions, and may be executed as code collectively executing on one or more processors (eg, executable instructions, one or more computer programs, or one or more applications), implemented in hardware, or a combination thereof. The computer program includes a plurality of instructions executable by one or more processors.

Further, the methods may be implemented in any type of computing platform operably connected to a suitable, including but not limited to personal computer, minicomputer, mainframe, workstation, network or distributed computing environment, stand-alone or integrated computer platform, or communicate with charged particle tools or other imaging devices, etc. Aspects of the invention may be implemented in machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, an optically read and/or written storage medium, RAM, ROM, etc., such that it can be read by a programmable computer, when a storage medium or device is read by a computer, it can be used to configure and operate the computer to perform the processes described herein. Furthermore, the machine-readable code, or portions thereof, may be transmitted over wired or wireless networks. The invention described in this embodiment includes these and other various types of non-transitory computer-readable storage media when such media includes instructions or programs that implement the steps described above in conjunction with a microprocessor or other data processor. The invention also includes the computer itself when programmed according to the methods and techniques described herein.

A computer program can be applied to input data to perform the functions described in this embodiment to transform the input data to generate output data for storage to non-volatile memory. The output information can also be applied to one or more output devices such as a display. In a preferred embodiment of the present invention, the transformed data represents the physical and tangible target terminals, including specific visual depictions of the physical and tangible target terminals produced on the display.

The above are only preferred embodiments of the present invention, and the present invention is not limited to the above-mentioned embodiments, as long as it achieves the technical effect of the present invention by the same means, all within the spirit and principle of the present invention, do Any modification, equivalent replacement, improvement, etc., should be included within the protection scope of the present invention. Various modifications and changes can be made to its technical solutions and/or implementations within the protection scope of the present invention.

Claims (10)

1. a training method of a product surface defect detection model, is characterized in that, comprises the following steps: Obtain an auto-encoding network; the auto-encoding network includes an encoder network and a decoder network; obtaining a sample image; at least a portion of the sample image includes product surface defects; training the auto-encoding network using the sample images; Obtain a generative adversarial network; the generative adversarial network includes a generator network and a discriminator network, and the discriminator network is constructed by the encoder network in the auto-encoding network; Obtain a real image and its mark; the real image includes product surface defects, and the mark is used to indicate the type of product surface defect included in the corresponding real image; Input the real image or the fake image generated by the generator network into the discriminator network, obtain the output result of the discriminator network, adjust the parameters of the discriminator network and/or the generator network, until The loss function of the discriminator network and/or the loss function of the generator network reaches the target value; The trained generative adversarial network is used as a product surface defect detection model. 2. The training method according to claim 1, wherein the step of adjusting the parameters of the discriminator network and/or the generator network specifically comprises: reconstructing sample images by the decoder network; Obtain the error of the decoder network in the reconstruction process; The encoder network is adjusted according to the error. 3. training method according to claim 2, is characterized in that, the loss function of described discriminator network is:

4 . The training method according to claim 1 , wherein the discriminator network is obtained by connecting the encoder network and a multi-classification network. 5 . 5. The training method according to claim 4, wherein the multi-classification network is a softmax network. 6. The training method according to claim 4, wherein the automatic coding network is a convolutional automatic coding network; the encoder network comprises a plurality of convolutional layers and a passthrough layer; the passthrough layer is used for The feature maps output by the shallower convolutional layers are downsized and concatenated with the feature maps output by the deeper convolutional layers. 7. The training method according to claim 6, wherein the deeper convolutional layer is the last convolutional layer in the encoder network. 8. The training method according to claim 4, wherein at least one of the convolutional layers in the encoder network performs a max-pooling operation. 9. A product surface defect detection method, is characterized in that, comprises the following steps: Obtain an image to be detected; the image to be detected includes the product surface; Input the image to be detected into the product surface defect detection model; the product surface defect detection model is trained by the training method described in any one of claims 1-8; The output result of the product surface defect detection model is acquired, and the product surface defect type is determined according to the output result. 10. A storage medium storing processor-executable instructions, wherein, when executed by the processor, the processor-executable instructions are used to execute the method according to any one of claims 1-9. method.

Images