Conveying belt defect detection method based on image defogging processing

Abstract

The invention discloses a conveyor belt defect detection method based on image defogging processing, and solves the problems of poor defogging effect and low defect detection accuracy in the prior art. The method comprises the steps that an original image is input into a defogging branch network and a backbone network, the defogging branch network carries out normalization and correction processing on the original image, and multi-scale defogging image features are obtained; the backbone network performs feature extraction on the original image to obtain multi-scale original image features; performing feature fusion on the multi-scale defogged image features and the multi-scale original image features to obtain fused features; and inputting the fused features into a neck network for processing, and inputting the processed features into a detection head for feature detection.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and specifically, relates to an image-based defect detection technology, and more specifically, relates to a conveyor belt defect detection method based on image defogging processing. Background Art

[0002] Conveyor belts are key material transport equipment in industry and are widely used in manufacturing, mining, agriculture, and logistics. However, long-term use can easily lead to defects in conveyor belts. If they are not detected and repaired in time, they may lead to production stagnation, economic losses, and even safety accidents. In addition, harsh environmental conditions such as haze will further increase the difficulty of detection and the risk of false detection. Therefore, establishing an efficient detection model under haze conditions is crucial to ensure the safe operation of conveyor belts and prevent accidents.

[0003] In recent years, the development of deep learning has enabled image processing methods to perform well in conveyor belt defect detection. In related technologies, there are two types of defect detection under haze conditions: one is the image preprocessing method, which uses a defogging network to process the original image, directly outputs a defogging image, and then performs feature extraction and detection based on the defogging image; the other is the model optimization method, which first uses a backbone network to extract features from the original image, and then combines defogging detection with target detection. This method requires the backbone network and the defogging detection network to form a codec structure, which has a complex structure. Both of these defect detection methods have the problem of low clarity of defogging images and easy loss of image detail information, resulting in poor conveyor belt defect detection results. Summary of the invention

[0004] One of the purposes of the present invention is to provide a conveyor belt defect detection method based on image defogging processing to solve the problems of poor defogging effect and low defect detection accuracy in the prior art.

[0005] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions:

[0006] A conveyor belt defect detection method based on image defogging processing, comprising:

[0007] The original image is input into the defogging branch network and the main network respectively, the defogging branch network performs normalization and correction processing on the original image to obtain multi-scale defogging image features; the main network performs feature extraction on the original image to obtain multi-scale original image features;

[0008] Performing feature fusion on the multi-scale dehazed image features and the multi-scale original image features to obtain fused features;

[0009] The fused features are input into the neck network for processing, and the processed features are input into the detection head for feature detection.

[0010] In some embodiments of the present application, the defogging branch network performs normalization and correction processing on the original image to obtain multi-scale defogging image features, including:

[0011] Performing convolution processing on the original image to obtain a primary feature map;

[0012] Performing convolution processing on the primary feature map to obtain a high-level feature map;

[0013] Calculate the mean and standard deviation of each pixel position of the primary feature map;

[0014] Normalizing the primary feature map according to the mean and the standard deviation to obtain a normalized feature map; correcting the high-level feature map according to the mean and the standard deviation to obtain a corrected feature map;

[0015] Splicing the normalized feature map and the corrected feature map to obtain a spliced ​​feature map;

[0016] The concatenated feature map is subjected to convolution processing and feature extraction fusion processing to obtain the multi-scale defogging image features.

[0017] In some embodiments of the present application, the primary feature map includes a first primary feature map and a second primary feature map, and the high-level feature map includes a first high-level feature map and a second high-level feature map;

[0018] The original image is convolved to obtain a primary feature map; the primary feature map is convolved to obtain a high-level feature map, specifically including:

[0019] Performing convolution processing on the original image through a first convolution module to obtain the first primary feature map;

[0020] Performing convolution processing on the first primary feature map through a second convolution module to obtain the second primary feature map;

[0021] splicing the first primary feature map and the second primary feature map, and performing convolution processing on the spliced ​​feature map through a third convolution module to obtain the first high-level feature map;

[0022] The second primary feature map and the first high-level feature map are spliced, and the spliced ​​feature map is convolved through a fourth convolution module to obtain the second high-level feature map.

[0023] In some embodiments of the present application, the normalized feature map includes a first normalized feature map and a second normalized feature map, and the corrected feature map includes a first corrected feature map and a second corrected feature map;

[0024] Calculating the mean and standard deviation of each pixel position of the primary feature map, and obtaining the normalized feature map and the corrected feature map according to the mean and the standard deviation, specifically includes:

[0025] Calculating the mean and standard deviation of each pixel position of the first primary feature map and the mean and standard deviation of each pixel position of the second primary feature map respectively;

[0026] The first primary feature map is normalized according to the mean and standard deviation of the first primary feature map to obtain the first normalized feature map; the first high-level feature map is corrected according to the mean and standard deviation of the first primary feature map to obtain the first corrected feature map; the second primary feature map is normalized according to the mean and standard deviation of the second primary feature map to obtain the second normalized feature map; the second high-level feature map is corrected according to the mean and standard deviation of the second primary feature map to obtain the second corrected feature map.

[0027] In some embodiments of the present application, the multi-scale dehazed image features and the multi-scale original image features are subjected to feature fusion to obtain fused features, including:

[0028] Adding the multi-scale dehazed image features and the multi-scale original image features to obtain mapping features;

[0029] The mapping features are average-pooled along the height direction and the width direction respectively to obtain two one-dimensional feature sequences;

[0030] Dividing the two one-dimensional feature sequences into multiple sub-features respectively, performing multi-scale depth-separable convolution processing on the multiple sub-features respectively, and processing the processed sub-features of the same dimension to obtain two significant features;

[0031] Multiplying the two significant features by the mapping feature element by element to obtain a spatial feature;

[0032] A multi-head self-attention mechanism is used to calculate the dependencies between channels for the spatial features to obtain channel features, average pooling is performed on the channel features, and channel attention is generated under the action of an activation function;

[0033] The channel attention is multiplied element-by-element by the spatial feature to obtain the fused feature.

[0034] In some embodiments of the present application, the processed sub-features of the same dimension are processed to obtain two significant features, including:

[0035] The processed sub-features of the same dimension are concatenated, group normalization is performed on the concatenated features, and then the salient features are generated under the action of an activation function.

[0036] In some embodiments of the present application, in the neck network, the fused features are dynamically upsampled to obtain a high-resolution feature map.

[0037] Another object of the present invention is to provide an electronic device, comprising a processor, a memory and a computer program stored in the memory, wherein the processor is configured to execute the computer program to implement the above-mentioned conveyor belt defect detection method based on image defogging processing.

[0038] Another object of the present invention is to provide a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the above-mentioned conveyor belt defect detection method based on image defogging is implemented.

[0039] Compared with the prior art, the advantages and positive effects of the present invention are:

[0040] The conveyor belt defect detection method based on image defogging proposed in the present invention sets a parallel defogging branch network and a backbone network, uses the defogging branch network to obtain multi-scale defogging image features from the original image, uses the backbone network to obtain multi-scale original image features from the original image, then fuses the multi-scale defogging image features and the multi-scale original image features, performs feature detection based on the fused features, and normalizes and corrects the original image through the defogging branch network, thereby enhancing the defogging effect, better retaining the image details, and improving the image clarity and contrast; then the multi-scale defogging image features obtained by the defogging branch network are also used as learning features, and are fused with the multi-scale original image features for subsequent processing, which can not only obtain image defogging information, but also retain the original image information, thereby enriching the learning information of the network, and further improving the accuracy and robustness of defect detection.

[0041] After reading the specific embodiments of the present invention in conjunction with the accompanying drawings, other features and advantages of the present invention will become more clear. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, a brief introduction will be given below to the drawings required for use in the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0043] Figure 1 It is a flow chart of an embodiment of a conveyor belt defect detection method based on image defogging processing proposed by the present invention;

[0044] Figure 2 A network structure diagram of an embodiment of a conveyor belt defect detection method based on image defogging processing proposed by the present invention;

[0045] Figure 3 A defogging branch network structure diagram in an embodiment of a conveyor belt defect detection method based on image defogging processing proposed by the present invention;

[0046] Figure 4 The collected image in the embodiment of the conveyor belt defect detection method based on image defogging processing proposed by the present invention;

[0047] Figure 5 for Figure 4 The image after adding fog to the part of the collected image;

[0048] Figure 6 This is a comparison chart of loss function results using different methods for defect detection;

[0049] Figure 7 This is another comparison chart of the test results using different methods for defect detection;

[0050] Figure 8 The present invention is a structural block diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0052] It should be noted that the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the fact that ordinary technicians in the field can implement it. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0053] Figure 1 and Figure 2 The flowchart and network structure diagram of an embodiment of a conveyor belt defect detection method based on image defogging processing proposed by the present invention are respectively shown.

[0054] See also Figure 1 and Figure 2 ,The implementation process of the conveyor belt defect detection method based on image defogging is as follows:

[0055] S11: Input the original image into the defogging branch network and the trunk network respectively to obtain multi-scale defogging image features and multi-scale original image features.

[0056] The original image refers to an image containing the conveyor belt on which defect detection is based, and may be an image of the conveyor belt directly captured by an image acquisition device, or an image obtained by preprocessing the image of the conveyor belt captured by the image acquisition device.

[0057] After the original image is obtained, the image features are extracted from the original image using a feature extraction network. In this embodiment, two feature extraction networks are provided in parallel, namely a defogging branch network and a trunk network.

[0058] The original image is input into the defogging branch network, which normalizes and corrects the original image to obtain multi-scale defogging image features. The defogging branch network can use a known network structure to normalize and correct the image and output multi-scale defogging image features. The multi-scale defogging image features obtained by the defogging branch network improve the clarity of the image in terms of color, texture and structure.

[0059] The original image is input into the backbone network, and the backbone network extracts features from the original image to obtain multi-scale original image features, which are image features without defogging. The multi-scale original image features that have not been defogged retain information such as the color, texture, and structure of the original image. The backbone network can use a known network structure to extract image features, which is not limited in this embodiment. For example, in some embodiments, the backbone network uses a compact inversion block (CIB), a large kernel convolution, and a local self-attention module (PSA) to achieve efficient extraction of image features. The CIB module extracts spatial and channel features through deep separable convolution, reduces computational costs, and maintains feature expression. Large kernel convolution expands the receptive field and enhances the global perception of large targets. The PSA module enhances global feature modeling and optimizes complex scenes and large target detection.

[0060] S12: Fusing the multi-scale dehazed image features and the multi-scale original image features to obtain fused features.

[0061] S13: Input the fused features into the neck network for processing, and input the processed features into the detection head for feature detection.

[0062] The process of processing the fused features using the neck network and the process of detecting features using the detection head can both be implemented using existing technologies, and this embodiment does not limit this.

[0063] The conveyor belt defect detection method of the above-mentioned embodiment sets a parallel defogging branch network and a backbone network as a feature extraction network, uses the defogging branch network to obtain multi-scale defogging image features from the original image, uses the backbone network to obtain multi-scale original image features from the original image, and then fuses the multi-scale defogging image features and the multi-scale original image features, performs feature detection based on the fused features, and normalizes and corrects the original image through the defogging branch network, thereby enhancing the defogging effect, better retaining the image details, and improving the image clarity and contrast; then the multi-scale defogging image features obtained by the defogging branch network are also used as learning features, and are fused with the multi-scale original image features for subsequent processing, which can not only obtain image defogging information, but also retain the original image information, thereby enriching the learning information of the network, and then helping to improve the accuracy and robustness of defect detection.

[0064] In some other embodiments, the defogging branch network performs normalization and correction processing on the original image to obtain multi-scale defogging image features, specifically including:

[0065] Perform convolution on the original image to obtain a primary feature map. Perform convolution on the primary feature map to obtain a high-level feature map.

[0066] Calculate the mean and standard deviation of each pixel position of the primary feature map. Specifically, the calculation formulas for the mean and standard deviation are as follows:

[0067]

[0068] In the formula, X b,c,h,w represents the primary feature map, μ b,h,w Represents the primary feature map X b,c,h,w The mean value, σ b,h,w Represents the primary feature map X b,c,h,w The standard deviation of , b represents the batch, c represents the number of channels, h represents the height of the feature map, w represents the width of the feature map, C 0 Represents the primary feature map X b,c,h,w ε is a constant used to adjust the standard deviation to avoid numerical instability when calculating the standard deviation.

[0069] Then, the primary feature map is normalized according to the mean and standard deviation to obtain the normalized feature map. Normalized feature map X ′ b,c,h,w It is expressed as: Among them, β is a constant used to control the translation amount of the data.

[0070] The primary feature map is a feature map obtained by directly processing the original image or after preliminary conversion of the original image. It is highly sensitive to local changes in the image and is particularly important for maintaining image details. By normalizing the primary feature map using the mean and standard deviation and normalizing each position separately, it helps to conduct more effective feature learning and abstraction in the future, and can ensure that the network can process all parts of the image evenly, enhance the network's processing capabilities for local features, and enable the model to better adapt to the feature differences in different regions when facing foggy images, providing a basis for subsequent feature correction and fusion, so as to improve the dehazing effect and the overall clarity of the image.

[0071] The high-level feature map is also corrected according to the mean and standard deviation to obtain the corrected feature map. Specifically, first, the standard deviation σ b,h,w The high-level feature map is scaled to make the overall distribution of the high-level feature map more uniform, which is convenient for the fusion of multi-scale information. Then, the mean μ b,h,w The scaled feature map is translated and adjusted to restore the numerical range of the feature map while retaining key spatial information and local details, thereby obtaining a corrected feature map.

[0072] The corrected feature map obtained through the above correction processing ensures that in the multi-scale feature fusion process, features of different scales can be fully utilized, effectively improving the feature expression ability and detail retention, while maintaining the stability and effectiveness of the network.

[0073] Finally, the normalized feature map and the corrected feature map are spliced ​​to obtain the spliced ​​feature map; the spliced ​​feature map is then subjected to convolution processing and feature extraction fusion processing to further optimize feature extraction and enhance dehazing effect, and obtain multi-scale dehazing image features.

[0074] The number of primary feature maps may be one or more; the number of advanced feature maps may be one or more, and the specific number may be determined according to actual needs.

[0075] Figure 3 The defogging branch network structure diagram in the embodiment of the conveyor belt defect detection method based on image defogging processing proposed by the present invention is shown. Figure 3 In the illustrated embodiment, the defogging branch network generates two primary feature maps and two high-level feature maps. Figure 3 , further describing in detail the process of performing defogging on the original image to obtain multi-scale defogging image features in this embodiment.

[0076] Get two primary feature maps:

[0077] The original image Input is first convolved through the first convolution module Conv1 to obtain the first primary feature map X 1 The first convolution module Conv1 is a 1×1 convolution module.

[0078] For the first primary feature map X 1 The second convolution module Conv2 performs convolution processing to obtain the second primary feature map X 2 The second convolution module Conv2 is a 3×3 convolution module.

[0079] Normalize the two primary feature maps to obtain two normalized feature maps:

[0080] For the first primary feature map X 1 Normalize ( Figure 3 PONO), calculate the first primary feature map X 1 The mean μ at each pixel position 1 and standard deviation σ 1 , according to the mean μ 1 and standard deviation σ 1 For the first primary feature map X 1 Perform normalization to obtain the first normalized feature map X ′ 1 .

[0081] For the second primary feature map X 2 Normalize ( Figure 3 PONO), calculate the second primary feature map X 2 The mean μ at each pixel position 2 and standard deviation σ 2 , according to the mean μ 2 and standard deviation σ 2 For the second primary feature map X 2 Perform normalization to obtain the second normalized feature map X ′ 2 .

[0082] Two high-level feature maps are obtained:

[0083] The first primary feature map X 1 and the second primary feature map X 2 To splice ( Figure 3 The concatenated feature map is convolved through the third convolution module Conv3 to obtain the first high-level feature map X 3 The third convolution module Conv3 is a 5×5 convolution module.

[0084] The second primary feature map X 2With the first high-level feature map X 3 The spliced ​​feature map is then convolved through the fourth convolution module Conv4 to obtain the second high-level feature map X 4 The fourth convolution module Conv3 is a 7×7 convolution module.

[0085] The two high-level feature maps are corrected to obtain two corrected feature maps:

[0086] According to the first primary feature map X 1 The standard deviation σ 1 and mean μ 1 Sequentially, the first high-level feature map X 3 Perform correction processing for scaling and translation ( Figure 3 MS), the first corrected feature map X is obtained ′ 3 .

[0087] According to the second primary feature map X 2 The standard deviation σ 2 and mean μ 2 Sequentially for the second high-level feature map X 4 Perform correction processing for scaling and translation ( Figure 3 MS), and obtain the second corrected feature map X ′ 4 .

[0088] Multi-feature map splicing processing:

[0089] The first normalized feature map X ′ 1 , the second normalized feature map X ′ 2 , the first corrected feature map X ′ 3 and the second corrected feature map X ′ 4 To splice ( Figure 3 Concat is represented in the figure to obtain the concatenated feature map, and then the concatenated feature map is convolved through the fifth convolution module Conv5 to obtain the feature map X 5 The fifth convolution module Conv5 is a 3×3 convolution module.

[0090] For feature map X 5 Then convolution and feature extraction fusion processing are performed to obtain the final multi-scale dehazed image features.

[0091] In some embodiments, in order to more effectively capture key features, an adaptive feature fusion method is used to fuse multi-scale dehazed image features and multi-scale original image features to obtain fused features.

[0092] The process of adaptive feature fusion is as follows:

[0093] First, the multi-scale dehazed image features and the multi-scale original image features are added to obtain the mapping features.

[0094] Then, the mapping features are average pooled along the height and width directions to obtain two one-dimensional feature sequences, which represent the overall information of different spatial dimensions. The two one-dimensional feature sequences are then divided into multiple sub-features, and the multiple sub-features are processed by multi-scale depth-separable convolution. The processed sub-features of the same dimension are spliced, group normalized, and activated to obtain two significant features. The two significant features are then element-wise multiplied with the mapping features to obtain spatial features.

[0095] In the process of obtaining spatial features, the one-dimensional feature sequence is divided into multiple sub-features for processing to achieve the purpose of extracting multi-scale spatial information, which not only reduces the computational complexity, but also enhances the expression of local features and achieves the unity of detail information and global information. By performing multi-scale deep separable convolution processing on multiple sub-features, spatial structure information of different scales can be effectively extracted.

[0096] Then, the multi-head self-attention mechanism is used to calculate the dependencies between channels for the spatial features, obtain the channel features, perform average pooling on the channel features, and generate channel attention under the action of the activation function.

[0097] In this process, convolution is used to generate query matrix, key matrix and value matrix, and the multi-head self-attention mechanism uses each matrix to calculate the dependency between channels, and the obtained channel features have high discrimination and stability. The specific implementation of the multi-head self-attention mechanism using the matrix to calculate the dependency between channels to obtain channel features can adopt the existing technology, which will not be described in more detail here.

[0098] Finally, the channel attention is multiplied element-by-element with the spatial features to complete the channel calibration and obtain the fused features.

[0099] In some other embodiments, in order to further retain the key information of small targets and improve the detection accuracy of small targets, when the fused features are input into the neck network for processing, the neck network dynamically upsamples the fused features to convert the low-resolution feature map into a high-resolution feature map. In the dynamic upsampling process, by dynamically generating the sampling point position, the details are accurately captured, the restoration capability and computational efficiency of the high-resolution feature map are improved, and the detection of small targets is effectively improved.

[0100] The technical effect of the conveyor belt defect detection method proposed by the present invention is further explained below through the analysis of the results of specific embodiments.

[0101] Collect multiple images containing conveyor belt defects and annotate them. The annotated content includes two types of typical defects: one is the tearing and cracking of the conveyor belt, and the other is the laser line abnormality. Among them, the laser line abnormality assists detection through different abnormal states of the laser line (such as breakpoints, disconnection, etc.), helping to reduce the false detection rate and missed detection rate. Some of the annotated collected images are as follows: Figure 4 As shown in (a)-(f).

[0102] The collected images are fogged by the atmospheric scattering model. Some of the labeled collected images after fogging are as follows: Figure 5 As shown in (a), (b) and (c).

[0103] Build the YOLOv10 model as the basic framework, and make three improvements on the basic framework.

[0104] The first improvement (Model 1): add a dehazing branch in parallel with the backbone network.

[0105] The second improvement (Model 2): ​​add a dehazing branch in parallel with the backbone network; use an adaptive feature fusion method for feature fusion.

[0106] The third improvement (Model 3): Add a dehazing branch and connect it in parallel with the backbone network; use an adaptive feature fusion method for feature fusion; and use a dynamic upsampling method in the neck network to upsample the fused features.

[0107] The fogged image dataset is divided into training set, validation set and test set. The training set is used to train the three improved models. Then, the validation set is used to perform defect detection on the three improved models after training, and various performance indicators of different models are compared. The performance indicators include precision (P), recall (R), multi-category average accuracy (mAP@0.5) with loss function threshold (Interception over Union, IoU) of 0.5, and model inference speed (Frames Per Second, FPS).

[0108] The comparison results of various performance indicators of different models are shown in Table 1.

[0109] Table 1 Comparison of performance indicators of three models

[0110]

[0111] Analyzing the comparison results in Table 1, the three models with defogging branches added all have high precision, recall rate, average accuracy rate, and high recognition accuracy rate for conveyor belt defect detection. Model 2 and Model 3 using the adaptive feature fusion method have a higher recall rate and average accuracy rate than Model 1 which does not use the adaptive feature fusion method, although the inference speed of the model is slower. Model 3, which uses the defogging branch, the adaptive feature fusion method, and the dynamic upsampling method, has the highest precision, recall rate, and average accuracy rate.

[0112] We also used the same fogged image dataset to perform defect detection using Model 3 and five other existing models, and compared the accuracy, recall, average precision, and inference speed of various models. The comparison results are shown in Table 2.

[0113] Table 2 Comparison results of performance indicators of six models

[0114]

[0115] The comparison results in Table 2 show that Model 3 outperforms other models in multiple performance indicators, with an accuracy of 90.5%, a recall rate of 88.1%, and an average accuracy of 93.3%. These three indicators are 3.1%, 3.3%, and 3.1% higher than those of the traditional YOLOv10s, respectively. The comparison results show that although the detection speed of Model 3 has slightly decreased, it has significantly reduced the false detection rate and missed detection rate while maintaining a high frame rate.

[0116] Figure 6 The figure shows the comparison of the loss function results of the model 3 constructed by the method of the present invention and the model YOLOv10s in the prior art. Figure 6 Curve a in is the loss function curve of YOLOv10s. Figure 6 The curve b in FIG. 1 is a loss function curve diagram of the method of the present invention. Figure 6 It can be seen that after 500 rounds of training, the method proposed in the present invention shows significant advantages in the bounding box regression loss curve, and the convergence speed, training stability and final loss value are significantly lower than the prior art YOLOv10s. Moreover, the loss value of the method proposed in the present invention drops rapidly in the early stage of training, and the fluctuation is small. It can quickly learn effective features, and the model has high robustness, which significantly reduces the training error. Therefore, the method proposed in the present invention performs better in multi-scale targets and complex environments.

[0117] The model 3 proposed in the present invention and the YOLOv10s of the prior art are used for testing on the test set. Table 3 shows the performance indicators of using the model 3 to detect different types of defects.

[0118] Table 3 Different defect detection performance indicators

[0119]

[0120] It can be seen from Table 3 that the accuracy, recall rate and average precision of model 3 for laser line defect detection are all greater than 95%.

[0121] Figure 7 Another comparison chart of the detection results of the model 3 constructed by the method of the present invention and the model YOLOv10s in the prior art is shown, that is, a comparison chart of the test results on the test set. Figure 7 (a) is the test result diagram using YOLOv10s. Figure 7 (b) is a test result diagram using the method of the present invention. In the test result diagram, the value after the category is the confidence level, which indicates the probability of detecting this type of defect. Tear indicates tear defect, Fissure indicates crack defect, and F law indicates laser line abnormality defect. Figure 7 It can be seen from the comparison diagram that the method proposed in the present invention significantly improves the detection accuracy of various defects under haze conditions and the recognition probability of small targets, and reduces the false detection rate and missed detection rate.

[0122] Figure 8 The structural block diagram of an embodiment of the electronic device of the present invention is shown. The electronic device includes a processor 81, a memory 82, and a computer program 821 stored in the memory 82. The processor 81 is configured to execute the computer program 821 to implement the conveyor belt defect detection method based on image defogging processing of the above embodiment and achieve the technical effects of the corresponding embodiment.

[0123] Other embodiments of the present invention further provide a computer storage medium having a computer program stored thereon. When the computer program is executed by a processor, the conveyor belt defect detection method based on image defogging of the above-mentioned embodiment is implemented, and the technical effects of the corresponding embodiments are achieved.

[0124] The above-mentioned computer storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. Computer storage medium can be any available storage medium that can be accessed by general or special-purpose computers.

[0125] In some embodiments, a computer storage medium is coupled to a processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor and the storage medium can also exist in a device as discrete components.

[0126] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for a person skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to replace some of the technical features therein by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions claimed to be protected by the present invention.

Claims

1. A conveyor belt defect detection method based on image defogging, characterized in that: The method comprises: The original image is input into the defogging branch network and the main network respectively, the defogging branch network performs normalization and correction processing on the original image to obtain multi-scale defogging image features; the main network performs feature extraction on the original image to obtain multi-scale original image features; Performing feature fusion on the multi-scale dehazed image features and the multi-scale original image features to obtain fused features; The fused features are input into the neck network for processing, and the processed features are input into the detection head for feature detection.

2. The conveyor belt defect detection method based on image defogging according to claim 1 is characterized in that: The defogging branch network performs normalization and correction processing on the original image to obtain multi-scale defogging image features, including: Performing convolution processing on the original image to obtain a primary feature map; Performing convolution processing on the primary feature map to obtain a high-level feature map; Calculate the mean and standard deviation of each pixel position of the primary feature map; Normalizing the primary feature map according to the mean and the standard deviation to obtain a normalized feature map; correcting the high-level feature map according to the mean and the standard deviation to obtain a corrected feature map; Splicing the normalized feature map and the corrected feature map to obtain a spliced ​​feature map; The concatenated feature map is subjected to convolution processing and feature extraction fusion processing to obtain the multi-scale defogging image features.

3. The conveyor belt defect detection method based on image defogging according to claim 2 is characterized in that: The primary feature map includes a first primary feature map and a second primary feature map, and the high-level feature map includes a first high-level feature map and a second high-level feature map; Performing convolution processing on the original image to obtain a primary feature map; The primary feature map is convolved to obtain a high-level feature map, specifically including: Performing convolution processing on the original image through a first convolution module to obtain the first primary feature map; Performing convolution processing on the first primary feature map through a second convolution module to obtain the second primary feature map; splicing the first primary feature map and the second primary feature map, and performing convolution processing on the spliced ​​feature map through a third convolution module to obtain the first high-level feature map; The second primary feature map and the first high-level feature map are spliced, and the spliced ​​feature map is convolved through a fourth convolution module to obtain the second high-level feature map.

4. The conveyor belt defect detection method based on image defogging processing according to claim 3 is characterized in that: The normalized feature map includes a first normalized feature map and a second normalized feature map, and the corrected feature map includes a first corrected feature map and a second corrected feature map; Calculating the mean and standard deviation of each pixel position of the primary feature map, and obtaining the normalized feature map and the corrected feature map according to the mean and the standard deviation, specifically includes: Calculating the mean and standard deviation of each pixel position of the first primary feature map and the mean and standard deviation of each pixel position of the second primary feature map respectively; The first primary feature map is normalized according to the mean and standard deviation of the first primary feature map to obtain the first normalized feature map; the first high-level feature map is corrected according to the mean and standard deviation of the first primary feature map to obtain the first corrected feature map; the second primary feature map is normalized according to the mean and standard deviation of the second primary feature map to obtain the second normalized feature map; the second high-level feature map is corrected according to the mean and standard deviation of the second primary feature map to obtain the second corrected feature map.

5. The conveyor belt defect detection method based on image defogging according to any one of claims 1 to 4, characterized in that: The multi-scale dehazed image features and the multi-scale original image features are subjected to feature fusion. Obtain fused features, including: Adding the multi-scale dehazed image features and the multi-scale original image features to obtain mapping features; The mapping features are average-pooled along the height direction and the width direction respectively to obtain two one-dimensional feature sequences; Dividing the two one-dimensional feature sequences into multiple sub-features respectively, performing multi-scale depth-separable convolution processing on the multiple sub-features respectively, and processing the processed sub-features of the same dimension to obtain two significant features; Multiplying the two significant features by the mapping feature element by element to obtain a spatial feature; A multi-head self-attention mechanism is used to calculate the dependencies between channels for the spatial features to obtain channel features, average pooling is performed on the channel features, and channel attention is generated under the action of an activation function; The channel attention is multiplied element-by-element by the spatial feature to obtain the fused feature.

6. The conveyor belt defect detection method based on image defogging according to claim 5 is characterized in that: The processed sub-features of the same dimension are processed to obtain two significant features, including: The processed sub-features of the same dimension are concatenated, group normalization is performed on the concatenated features, and then the salient features are generated under the action of an activation function.

7. The conveyor belt defect detection method based on image defogging according to any one of claims 1 to 4, characterized in that: In the neck network, the fused features are dynamically upsampled to obtain a high-resolution feature map.

8. An electronic device comprising a processor, a memory and a computer program stored in the memory, characterized in that: The processor is configured to execute the computer program to implement the conveyor belt defect detection method based on image defogging as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the conveyor belt defect detection method based on image defogging described in any one of claims 1 to 7 is implemented.

Images