CN117635585A - Texture surface defect detection method based on teacher-student network - Google Patents

Abstract

The invention belongs to the field of image processing, and particularly relates to a texture surface defect detection method based on a teacher-student network, which is realized by adopting a defect detection model which is trained based on texture surface defect images and comprises a student network and a double-branch decoding module, and comprises the following steps: respectively encoding images to be detected by adopting a pre-training teacher network and a student network; carrying out semantic segmentation and feature recovery on the deepest scale features in the multi-scale features obtained by the student network by adopting a double-branch decoding module, and correspondingly obtaining a defect segmentation image and pseudo normal features with the same multi-scale as the multi-scale features obtained by the teacher network; calculating cosine similarity between the pseudo-normal features under each scale and features obtained by a teacher network to construct a global anomaly score map; and fusing the defect segmentation image and the global anomaly score image to obtain a detection result image, and completing defect detection. The invention can realize accurate detection of various texture defects in complex industrial environments.

Description

A texture surface defect detection method based on teacher-student network

Technical field

The invention belongs to the field of image processing, and more specifically, relates to a texture surface defect detection method based on a teacher-student network.

Background technique

Modern industrial production usually involves a high degree of automation and high-speed production, and the production process is becoming more and more complex, resulting in the inevitable defects or defects on the surface of many products, seriously affecting the yield rate and production efficiency of the production line. With the continuous advancement of computer image processing technology, texture surface defect detection methods based on machine vision continue to emerge. However, the normal areas and defective areas on the surface of most textured products usually have low contrast characteristics. Specifically, there is no obvious transition area, which makes it difficult to clearly image and distinguish clear contour boundaries. The texture background is seriously interfered, resulting in detection results. Not good. Moreover, there are different degrees of interference between defect areas and texture areas in the case of irregular textures and regular textures, which requires the detection technology to have good generalization. Due to the complex diversity of textures and defects, as well as the unpredictable and large-scale changes of defects, the robust detection of textured surface defects has always been a very challenging task.

As major manufacturing companies pay more and more attention to the detection of product surface defects, the traditional method of identifying defects by human eye observation has been gradually replaced by non-contact, high-speed, and high-precision methods due to its great subjectivity, instability, and low efficiency. It is replaced by modern automatic optical inspection (AOI) equipment. The theoretical basis of AOI technology comes from machine vision, which mainly includes four parts: visual imaging, visual positioning, visual detection, and visual classification. The core of which is the detection algorithm of the visual detection part. . The upper limit of the detection effect of AOI equipment usually depends on the performance of the detection algorithm. Therefore, high-performance texture surface defect detection algorithms are studied to achieve high-precision and highly robust detection of various texture surface defects, which is important for improving the quality and efficiency of automated industrial production. There's important meaning.

A large number of texture surface defect detection algorithms have been proposed, but these algorithms are difficult to achieve high-precision and efficient detection performance in actual production scenarios. Methods based on image reconstruction try to reconstruct the defective image into a normal image, and then calculate the residual between the original image and the reconstructed image to detect defects. However, this method is prone to reconstructing the defects well during the reconstruction process, resulting in leakage. Check. The method based on embedding similarity needs to continuously calculate and store the embedding vector of the normal image during the training process, and then calculate the distance between the embedding vector of the image to be tested and the stored normal embedding vector during the test phase. The distance between the embedding vector and the stored normal embedding vector is judged. As a drawback, the detection results of this method are not only rough but also require additional storage overhead. The storage overhead will increase linearly with the growth of the training set. At the same time, due to the need to calculate the distance between a large number of vectors, the calculation process of this algorithm is extremely time-consuming. , It is difficult to meet the real-time needs in industrial production.

Therefore, in order to promote the improvement of the production process and increase the yield rate, it is necessary to propose a robust texture surface defect detection algorithm that can achieve excellent detection results in the face of complex and changeable production environments and texture surface defects.

Contents of the invention

In view of the shortcomings and improvement needs of the existing technology, the present invention provides a texture surface defect detection method based on a teacher-student network. Its purpose is to propose a robust texture surface defect detection algorithm for various tasks in complex industrial environments. Texture-like defects can be accurately detected.

In order to achieve the above object, according to one aspect of the present invention, a texture surface defect detection method based on a teacher-student network is provided, using a texture surface defect detection method including a student network and a dual-branch decoding module trained based on texture surface defect images. Model implementation, including:

The pre-trained teacher network and the student network are used to encode the texture surface image to be detected respectively to obtain corresponding multi-scale features;

The dual-branch decoding module is used to perform semantic segmentation and feature recovery on the deepest scale features among the multi-scale features obtained by the student network, corresponding to the defect segmentation image and the multi-scale features obtained by the teacher network. Pseudo-normal features with the same multiple scales;

Calculate the cosine similarity between each scale feature in the multi-scale pseudo-normal features and the corresponding scale feature in the multi-scale features obtained by the teacher network to obtain the corresponding abnormal score map; convert the abnormal score map of each scale Upsample to the size of the texture surface image to be detected and multiply it pixel by pixel to obtain the global anomaly score map;

The defect segmentation image and the global anomaly score map are fused to obtain a detection result image to complete texture surface defect detection.

Furthermore, the texture surface defect image is an artificial defect image in the training stage, which is constructed in the following way:

A normal image of a textured surface is enhanced; a noise image generated by Perlin noise is thresholded and binarized to obtain an artificial defect mask image; another normal image of a textured surface is fused with the enhanced image The normal image of the textured surface and the artificial defect mask image are used to obtain the artificial defect image.

Further, the loss functions used in the training process of the texture surface defect detection model include: pixel-level contrast decoupled distillation loss, semantic segmentation loss, and feature recovery loss;

Among them, the pixel-level contrast decoupling distillation loss is used to explicitly guide the student network to distinguish between normal feature encoding and defect feature encoding in the image. The construction method is: downsampling the artificial defect mask image to obtain the texture and defect features; according to the semantic labels, the texture and defect features in the artificial defect feature embedding are decoupled and separated, and a normal feature vector set and a defect feature vector set are obtained. Based on the normal feature vector set and Defect feature vector set, calculate pixel-level contrast decoupling distillation loss to improve the similarity between the normal feature vector set and normal feature embedding and reduce the similarity between the defect feature vector set and normal feature embedding; so The artificial defect feature embedding is the deepest scale artificial defect feature among the multi-scale artificial defect features obtained by encoding the artificial defect image by the student network; the normal feature embedding is: the other image is encoded by the pre-trained teacher network The texture surface normal image is encoded, and the deepest scale feature among the encoded multi-scale normal features is fine-tuned in the spatial dimension by a linear mapper;

The semantic segmentation loss is used to supervise that the defect segmentation image output by the semantic segmentation branch in the dual-branch decoding module is close to the artificial defect mask image;

The feature recovery loss is used to supervise that the multi-scale pseudo-normal features output by the feature recovery branch in the dual-branch decoding module are close to the multi-scale normal features.

Further, the pixel-level contrast decoupling distillation loss is expressed as:

In the formula, L _intra represents the pixel-level contrast decoupling distillation loss, VN represents the number of normal feature vectors in the normal feature vector set, represents the i-th normal feature vector, VD represents the number of defect feature vectors in the defect feature vector set, /> Represents the jth defect feature vector,/> Represents the normal feature embedding with/> Feature vectors located at the same spatial position,/> Represents the normal feature embedding with/> Feature vectors located at the same spatial position, τ represents the temperature parameter used to smooth the data distribution, and sim(·) represents cosine similarity.

Further, the pixel-level contrast decoupling distillation loss is expressed as:

In the formula, L _c represents the pixel-level contrast decoupling distillation loss, Represents another random normal feature embedding of the same batch with/> Feature vectors located at the same spatial position,/> Represents the other normal feature embedding with/> Eigenvectors located at the same spatial location.

Further, the feature recovery loss is specifically a global feature mask-aware distillation loss, and its construction method is:

For each scale feature in the multi-scale normal features, the semantic tags to be trained together with the detection model are used for mask-aware distillation to generate a first feature-level mask corresponding to the scale; at the same time, each scale of the multi-scale pseudo-normal features is The feature uses the trainable semantic mark to perform mask-aware distillation to obtain a second feature-level mask corresponding to the scale; the first feature-level mask and the second feature-level mask at each scale are fused to obtain the Global perceptual mask at scale;

Sampling dense supervision method, calculating the L2 distance in the spatial dimension between normal features and pseudo-normal features at each scale, used to combine with the global perceptual mask at that scale, and calculate the feature mask perceptual distillation loss at this scale ;Synthesize the feature mask-aware distillation loss at all scales to obtain the global feature mask-aware distillation loss.

Furthermore, the feature mask-aware distillation loss at each scale is expressed as:

In the formula, _Hk represents the feature height at the kth scale, _Wk represents the feature width at the kth scale, ^Mk represents the global sensing mask at the kth scale, and ^Dk represents the L2 distance at the kth scale. .

Further, the semantic segmentation loss is expressed as:

In the formula, represents the calculation expectation, ||·|| ₁ represents the L1 norm, I _m represents the artificial defect mask image, I _s represents the defect segmentation image, and I _d represents the artificial defect image.

The present invention also provides a computer-readable storage medium. The computer-readable storage medium includes a stored computer program, wherein when the computer program is run by a processor, the device where the storage medium is located is controlled to execute the texture as described above. Surface defect detection methods.

Generally speaking, through the above technical solutions conceived by the present invention, the following beneficial effects can be achieved:

(1) The texture surface defect detection method proposed by the present invention is implemented by combining a pre-trained teacher network and a constructed texture surface defect detection model including a student network and a dual-branch decoding module. The dual-branch decoding module can achieve feature recovery and semantics segmentation. First, the student network is used to encode the image to be tested to obtain multi-scale features. The dual-branch decoding module performs semantic segmentation and feature recovery on the deepest scale features of the multi-scale features. During this period, the teacher network will also encode the image to be tested to obtain another A multi-scale feature. The number of pseudo-normal feature scales obtained by the above feature recovery needs to be the same as the number of feature scales obtained by the teacher network, so as to calculate the similarity between the pseudo features and the features obtained by the teacher network at each scale to construct global anomalies. Score map, this process makes full use of the knowledge of the pre-trained teacher network to generalize to the texture surface defect detection task, and combines the prior knowledge of the teacher network to achieve refined defect segmentation; further, combines the defect segmentation image and global anomalies obtained by semantic segmentation Score map, as the final detection result, the defect segmentation image result obtained by the semantic segmentation branch is relatively refined, suitable for detecting structural defects, and the outline is relatively clear, while the global anomaly score map obtained based on the feature recovery branch has good robustness and is suitable for detection. Therefore, the present invention proposes to integrate the results of these two branches, which can greatly improve the detection accuracy of various texture surface defects and has high application value.

(2) When training the texture surface defect detection model, the present invention introduces pixel-level contrast decoupling distillation loss, explicitly drives the network to learn differentiated representations of these two types of features during the training process, and guides the student network to identify normal features. Effective decoupling and separation from defect features can effectively enhance the discriminative ability of the model and further improve detection performance.

(3) When training the texture surface defect detection model, the present invention also introduces the global feature mask perceptual distillation loss, combined with the normal features of each scale obtained by the teacher network coding, to promote the transfer of effective knowledge by the feature recovery branch in the dual-branch decoding module. This allows the network to pay more attention to valuable prototype prior feature information during feature recovery, and comprehensively perceive the contextual dependencies between multi-level features, thereby increasing the detection rate and reducing the over-detection rate.

Description of drawings

Figure 1 is a flow chart of a texture surface defect detection method based on a teacher-student network provided by an embodiment of the present invention;

Figure 2 is a schematic diagram of the training process of the textured surface defect detection model provided by an embodiment of the present invention;

Figure 3 is a schematic diagram of the implementation of the sensing mask provided by the embodiment of the present invention;

Figure 4 is a schematic diagram of the feature comparison and decoupling effect provided by the embodiment of the present invention;

Figure 5 is a schematic diagram of the defect detection effect provided by the embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

Embodiment 1

A texture surface defect detection method based on a teacher-student network is implemented using a texture surface defect detection model including a student network and a dual-branch decoding module trained based on texture surface defect images, as shown in Figure 1, including:

The pre-trained teacher network and the above-mentioned student network are used to encode the texture surface image to be detected respectively to obtain their corresponding multi-scale features;

The above-mentioned dual-branch decoding module is used to perform semantic segmentation and feature recovery on the deepest scale features among the multi-scale features obtained by the student network. The corresponding defect segmentation image is obtained and has the same characteristics as the multi-scale features obtained by the teacher network. Pseudo-normal characteristics of scale;

Calculate the cosine similarity between each scale feature in the multi-scale pseudo-normal features and the corresponding scale feature in the multi-scale features obtained by the teacher network to obtain the corresponding anomaly score map; upsample the anomaly score map at each scale to the desired Detect the texture surface image size and multiply it pixel by pixel to obtain the global anomaly score map;

The defect segmentation image and the global anomaly score map are fused to obtain the detection result image to complete texture surface defect detection.

As a preferred implementation, the above-mentioned texture surface defect image is an artificial defect image in the training stage, which is constructed in the following manner:

A normal image of a textured surface is enhanced; a noise image generated by Perlin noise is thresholded and binarized to obtain an artificial defect mask image; another normal image of a textured surface is fused with the enhanced image The normal image of the textured surface and the artificial defect mask image are used to obtain the artificial defect image.

Specifically, taking the generation process of a single artificial defect image as an example, first, two normal images _It and _In are randomly selected without replacement. Then, a data enhancement transformation set is constructed: gamma correction, brightness transformation, hue transformation, overexposure enhancement, pixel inversion, Gaussian blur, elastic transformation, and as an option, three transformations are randomly selected from this set to process I _t , the enhanced image I _u is obtained:

I _u = f _aug (I _t );

Among them, I _u , _It ∈R ^W×H×C , W, H, and C represent the width, height, and channel number of the image respectively. For example, in this embodiment, they are selected as 256, 256, and 3 respectively; f _aug ( ·) represents a mapping function composed of three random transformations.

Next, the noise image I _p generated by Perlin noise is thresholded and binarized to obtain the artificial defect mask image _Im :

I _m = f _tb (I _p );

Among them, I _m , Illustratively, in this embodiment, C _m is selected as 1; f _tb (·) represents the threshold binarization operation.

Finally, the artificial defect image I _d is obtained by merging the images I _u , _In and _Im according to the following formula:

I _d =I _n ⊙(1-I _m )+I _u ⊙I _m ;

Among them, I _d , I _n ∈R ^W×H×C ; ⊙ represents the pixel-by-pixel multiplication operation, that is, the Hadamard product.

Furthermore, the training process of the above textured surface defect detection model can be:

The artificial defect images are encoded using the student network to be trained to obtain multi-scale artificial defect features, and the deepest scale artificial defect features of the multi-scale artificial defect features are embedded as the corresponding artificial defect features;

Using the dual-branch decoding module to be trained, semantic segmentation and feature recovery are performed on the above-mentioned artificial defect feature embeddings, corresponding to the defect segmentation image and multi-scale pseudo-normal features with the same number of scales as the multi-scale normal features; the multi-scale normal features It is obtained by encoding the above-mentioned texture surface normal image I _n by the pre-trained teacher network;

Based on the loss function, the parameters of the to-be-trained student network and the dual-branch decoding module are optimized, and the above process is repeated until convergence, and the texture surface defect detection model is obtained.

In this embodiment, the above-mentioned teacher network and student network as a whole can be regarded as a feature extraction mapping module. As shown in Figure 2, during the training process of the texture surface defect detection model, the feature extraction mapping module first encodes the texture surface normal image I _n and the corresponding artificial defect image I _d according to the following formula:

in, and/> represent the multi-scale features encoding the normal image I _n and the corresponding artificial defect image I _d respectively, /> Represents the three-dimensional features of the kth layer. N _n and N _d respectively represent the number of respective feature levels, that is, the number of scales. As an example, they are selected as 2 and 3 respectively; f _T (·) and θ _T respectively represent the feature extraction mapping module. Functions and parameters of the teacher network. The teacher network is pre-trained and the parameters are frozen and not updated. For example, ResNet18 can be used as the teacher network; f _Se (·) and θ _Se respectively represent the functions and parameters of the student network in the extraction mapping module.

Considering the domain differences between the pre-training data set of the teacher network and the target data set of the algorithm of the present invention, the feature extraction mapping module uses a linear mapper to map the deepest scale features among the multi-scale normal features. Fine-tuning is performed in the spatial dimension to obtain normal feature embedding Z _n :

in, As an example, W _l , H _l , and C _l are 32, 32, and 128 respectively; f _pt (·) and θ _pt respectively represent the functions and parameters of the above linear mapper. Z _n can be regarded as a set composed of a series of eigenvectors

As a preferred implementation, the loss functions used in the training process of the above-mentioned texture surface defect detection model include: pixel-level contrast decoupling distillation loss, semantic segmentation loss, and feature recovery loss.

Among them, the above-mentioned pixel-level contrast decoupled distillation loss is used to explicitly guide the student network to distinguish between normal feature encoding and defective feature encoding in the image, as shown in Figure 2. Its construction method is: under the above artificial defect mask image Sampling, the semantic labels of texture and defect features are obtained; according to the above semantic labels, the texture and defect features in the artificial defect feature embedding are decoupled and separated, and a normal feature vector set and a defect feature vector set are obtained. Based on the normal features Vector set and defect feature vector set, calculate pixel-level contrast decoupling distillation loss to improve the similarity between the normal feature vector set and the deepest scale feature in the above-mentioned multi-scale normal features and reduce the similarity between the above-mentioned defect feature vector set and the above-mentioned multi-scale The similarity between the deepest scale features in the normal features; the above-mentioned artificial defect feature embedding is the deepest scale artificial defect feature among the multi-scale artificial defect features obtained by encoding the artificial defect image by the student network.

The construction process of the above pixel-level contrast decoupling distillation loss can be called the contrast decoupling distillation module. The contrast decoupling distillation module embeds the above artificial defect features. (i.e., the deepest-scale artificial defect features of multi-scale artificial defect features) are regarded as a series of vectors in the spatial dimension/> The formed feature embedding Z _d , Z _d and Z _n are strictly aligned in each dimension (including the spatial dimension and the channel dimension), and the feature embedding Z _n of the normal image is used to guide the relationship between the normal features and defect features in Z _d Contrast decoupled distillation.

First, the artificial defect mask image I _m is downsampled to obtain the semantic labels I _l of two types of features (texture and defects):

I _l = f _ds (I _m );

in, f _ds (·) represents the downsampling operation.

Then, according to the label, the collection Classify the texture features and defect features (i.e. decoupling and separation):

in, and/> Represent normal feature vectors and defect feature vectors respectively, VN and VD respectively represent the number of two types of features in their respective sets.

Secondly, for subsequent convenience of expression, the cosine similarity between the two feature vectors will be calculated, expressed as the symbol sim(v _i , v _j ):

Among them, · represents the inner product, ||·|| ₂ represents the L2 norm.

Next, the pixel-level contrast decoupling distillation loss L _intra is calculated according to the following formula:

Among them, τ represents the temperature parameter used to smooth the data distribution, which can be set to 0.1 as an example. Represents the normal feature embedding with/> Feature vectors located at the same spatial position,/> Represents the normal feature embedding with/> Eigenvectors located at the same spatial location.

As a further preferred implementation, in order to enhance the robustness of the algorithm, artificial disturbances are introduced into the normal feature embeddings Z _n of the same batch, and randomly disrupted in the batch dimension to form a non-aligned vector set. Represents the feature vector in another random normal feature embedding of the same batch. On this basis, the contrast decoupling distillation _loss L inter between global images is calculated in the same form as calculating L _intra :

Finally, the contrastive decoupling distillation loss L _c composed of the two parts L _intra and L _inter as a whole:

In the embodiment, L _c gives equal weight to the two parts L _intra and L _inter , indicating that the confidence level for both parts is the same. As shown in Figure 4, the contrastive decoupling distillation loss L _c can provide effective constraints to promote class separability between features.

The dual-branch decoding module is used for semantic segmentation and feature recovery. The semantic segmentation represents:

I _s = f _Ss (Z _d ; θ _Ss );

Among them, I _s ∈R ^W×H×1 represents the defect segmentation image, f _Ss (·) and θ _Ss represent the functions and parameters of the semantic segmentation branch respectively. The above semantic segmentation loss is used to supervise the defect segmentation image output by the semantic segmentation branch in the dual-branch decoding module to be close to the artificial defect mask image. The segmentation loss L _s is calculated according to the following formula:

in, represents the calculation expectation, ||·|| ₁ represents the L1 norm, I _m represents the artificial defect mask image, I _s represents the defect segmentation image, and I _d represents the artificial defect image.

The above feature recovery is expressed as: Among them,/> Represents the multi-level pseudo-normal features recovered by the network, N _r represents the number of pseudo-normal feature levels, the value is the same as N _n , and the value is 2 for example; f _Sr (·) and θ _Sr respectively represent the function sum of the feature recovery branch parameter.

The above-mentioned feature recovery loss is used to supervise the multi-scale pseudo-normal features output by the feature recovery branch in the dual-branch decoding module to be close to the multi-scale normal features output by the teacher network. It can be used as a preferred implementation. The feature recovery loss is specifically a global feature mask. Perceptual distillation loss, whose construction can be implemented by the mask-aware distillation module, as shown in Figure 2, the mask-aware distillation module first introduces trainable semantic tags Generate feature-level masks, as shown in Figure 3. The formulaic calculation form of the perceptual mask at each scale is as follows:

in, and/> respectively represent the feature-level masks generated after semantic labeling processes normal features and restored pseudo-normal features, M ^k represents the global perceptual mask learned by the network; f _mg (·) and θ _mg respectively represent the mask-aware distillation Functions and parameters.

Then, a dense supervision method is used to calculate the L2 distance in the spatial dimension between the normal features and the restored pseudo-normal features at each scale:

Next, the feature mask-aware distillation loss for each scale is calculated as:

Finally, the global mask-aware distillation loss L _m is obtained by integrating multi-scale losses:

Among them, K represents the number of feature levels (that is, the number of scales) used in the algorithm, and the example is 2. L _m can effectively constrain the model to pay more attention to important contextual information distillation and transfer during the training process.

When training the texture surface defect detection model, the overall training loss L can be calculated based on the contrast decoupling distillation loss L _c , semantic segmentation loss L _s , and mask-aware distillation loss L _m obtained above, and the entire model can be trained and iterated accordingly. To convergence, the texture surface defect detection model is obtained. The overall loss of the algorithm during the training phase is calculated as follows:

L＝λ _c L _c +λ _s L _s +λ _m L _m ;

Among them, λ _c, λ _s , and λ _m represent adjustable hyperparameters used to balance the loss term, and are exemplary set to 0.1, 10, and 100 respectively.

After training and obtaining the texture surface defect detection model, an image to be tested can be input during the testing phase, and the features will be extracted through the feature extraction mapping module. The teacher network will output multi-level features. The features output by the student network will be processed by the dual-branch decoding module to obtain the defect segmentation image I _s and multi-level recovery features/> Exemplary N _n and N _r are both 2. By calculating the cosine similarity between the restored features at each level and the teacher network output features, the anomaly score map of each level is obtained:

These anomaly score maps are upsampled to the input image size, and then multiplied pixel by pixel to obtain the global anomaly score map I _a :

Among them, I _a ∈R ^W×H×1 , U(·) represents the interpolation upsampling operation.

Finally, the defect segmentation image I _s is fused with the anomaly score map I _a to obtain the detection result image I _r :

I _r =G(λ _f I _a +(1-λ _f )I _s ·max(I _a ));

Among them, I _r ∈ ^{R W × H × 1} , G represents Gaussian denoising, 0 ≤ λ _f ≤ 1 is used to adjust the weight ratio of the two items, and the exemplary setting is 0.5, which represents the defect segmentation image and the abnormal score image. The confidence level is the same; max(I _a ) represents the maximum value in the anomaly score map I _a .

Using the detection method provided in this embodiment, the detection effect of various texture surface defects is shown in Figure 5, in which the first row is the test image, the second row is the true value image, and the third row is the detection result in the test image. In the heat map above, the fourth row is the detection result image after binarization. It can be seen from the figure that the texture surface defect detection method proposed in this embodiment can achieve high-precision detection of various texture surface defects of different scales, different shapes, different contrasts and unknown types without using any real defect images. High-efficiency detection can effectively improve the quality and efficiency of automated industrial production.

Embodiment 2

A computer-readable storage medium, the computer-readable storage medium includes a stored computer program, wherein when the computer program is run by a processor, the device where the storage medium is located is controlled to execute one of the methods described in Embodiment 1 above. Textured surface defect detection methods.

The relevant technical solutions are the same as those in Embodiment 1 and will not be described again here.

It is easy for those skilled in the art to understand that the above descriptions are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements, etc., made within the spirit and principles of the present invention, All should be included in the protection scope of the present invention.

Claims (9)

1. A texture surface defect detection method based on a teacher-student network, which is characterized in that it is implemented using a texture surface defect detection model including a student network and a dual-branch decoding module trained based on texture surface defect images, including: The pre-trained teacher network and the student network are used to encode the texture surface image to be detected respectively to obtain corresponding multi-scale features; The dual-branch decoding module is used to perform semantic segmentation and feature recovery on the deepest scale features among the multi-scale features obtained by the student network, corresponding to the defect segmentation image and the multi-scale features obtained by the teacher network. Pseudo-normal features with the same multiple scales; Calculate the cosine similarity between each scale feature in the multi-scale pseudo-normal features and the corresponding scale feature in the multi-scale features obtained by the teacher network to obtain the corresponding abnormal score map; convert the abnormal score map of each scale Upsample to the size of the texture surface image to be detected and multiply it pixel by pixel to obtain a global anomaly score map; The defect segmentation image and the global anomaly score map are fused to obtain a detection result image to complete texture surface defect detection. 2. The texture surface defect detection method according to claim 1, characterized in that the texture surface defect image is an artificial defect image in the training stage, which is constructed in the following manner: A normal image of a textured surface is enhanced; a noise image generated by Perlin noise is thresholded and binarized to obtain an artificial defect mask image; another normal image of a textured surface is fused with the enhanced image The normal image of the textured surface and the artificial defect mask image are used to obtain the artificial defect image. 3. The texture surface defect detection method according to claim 2, characterized in that the loss function used in the training process of the texture surface defect detection model includes: pixel-level contrast decoupling distillation loss, semantic segmentation loss, and feature recovery loss; Among them, the pixel-level contrast decoupling distillation loss is used to explicitly guide the student network to distinguish between normal feature encoding and defect feature encoding in the image. The construction method is: downsampling the artificial defect mask image to obtain the texture and defect features; according to the semantic labels, the texture and defect features in the artificial defect feature embedding are decoupled and separated, and a normal feature vector set and a defect feature vector set are obtained. Based on the normal feature vector set and Defect feature vector set, calculate pixel-level contrast decoupling distillation loss to improve the similarity between the normal feature vector set and normal feature embedding and reduce the similarity between the defect feature vector set and normal feature embedding; so The artificial defect feature embedding is the deepest scale artificial defect feature among the multi-scale artificial defect features obtained by encoding the artificial defect image by the student network; the normal feature embedding is: the other image is encoded by the pre-trained teacher network The textured surface normal image is encoded, and the deepest scale feature among the encoded multi-scale normal features is fine-tuned in the spatial dimension by a linear mapper; The semantic segmentation loss is used to supervise that the defect segmentation image output by the semantic segmentation branch in the dual-branch decoding module is close to the artificial defect mask image; The feature recovery loss is used to supervise that the multi-scale pseudo-normal features output by the feature recovery branch in the dual-branch decoding module are close to the multi-scale normal features. 4. The texture surface defect detection method according to claim 3, characterized in that the pixel-level contrast decoupling distillation loss is expressed as: In the formula, L _intra represents the pixel-level contrast decoupling distillation loss, VN represents the number of normal feature vectors in the normal feature vector set, represents the i-th normal feature vector, VD represents the number of defect feature vectors in the defect feature vector set, /> Represents the jth defect feature vector,/> Represents the normal feature embedding with/> Feature vectors located at the same spatial position,/> Represents the normal feature embedding with/> Feature vectors located at the same spatial position, τ represents the temperature parameter used to smooth the data distribution, and sim(·) represents cosine similarity. 5. The texture surface defect detection method according to claim 3, characterized in that the pixel-level contrast decoupling distillation loss is expressed as: In the formula, L _c represents the pixel-level contrast decoupling distillation loss, Represents another random normal feature embedding of the same batch with/> Feature vectors located at the same spatial position,/> Represents the other normal feature embedding with/> Eigenvectors located at the same spatial location. 6. The texture surface defect detection method according to claim 3, characterized in that the feature recovery loss is specifically a global feature mask-aware distillation loss, and its construction method is: For each scale feature in the multi-scale normal features, the semantic tags to be trained together with the detection model are used for mask-aware distillation to generate a first feature-level mask corresponding to the scale; at the same time, each scale of the multi-scale pseudo-normal features is The feature uses the trainable semantic mark to perform mask-aware distillation to obtain a second feature-level mask corresponding to the scale; the first feature-level mask and the second feature-level mask at each scale are fused to obtain the Global perceptual mask at scale; Sampling dense supervision method, calculating the L2 distance in the spatial dimension between normal features and pseudo-normal features at each scale, used to combine with the global perceptual mask at that scale, and calculate the feature mask perceptual distillation loss at this scale ;Synthesize the feature mask-aware distillation loss at all scales to obtain the global feature mask-aware distillation loss. 7. The texture surface defect detection method according to claim 6, characterized in that the feature mask perceptual distillation loss at each scale is expressed as: In the formula, _Hk represents the feature height at the kth scale, _Wk represents the feature width at the kth scale, ^Mk represents the global sensing mask at the kth scale, and ^Dk represents the L2 distance at the kth scale. . 8. The texture surface defect detection method according to claim 3, characterized in that the semantic segmentation loss is expressed as: In the formula, represents the calculation expectation, ||·|| ₁ represents the L1 norm, I _m represents the artificial defect mask image, I _s represents the defect segmentation image, and I _d represents the artificial defect image. 9. A computer-readable storage medium, characterized in that the computer-readable storage medium includes a stored computer program, wherein when the computer program is run by a processor, the device where the storage medium is located is controlled to execute as claimed in claim The texture surface defect detection method described in any one of 1 to 8.