System and method for detecting microdefects on a strip surface

Abstract

The present application discloses a strip steel surface micro defect detection system and method, and relates to the technical field of defect detection, which comprises the following steps: obtaining an original image from a strip steel production line and constructing a database for strip steel surface micro defect detection; inputting a micro defect picture into a backbone feature extraction network to extract image features; according to the internal priori of production, i.e., the vertical direction homomorphism and the horizontal direction difference of micro defects, using an orthogonal context attention module to simultaneously enhance the context detail information from the vertical and horizontal directions, thereby enhancing the feature representation of small targets; using a cross adaptive aggregation module to adopt cross-layer semantic priori in the feature fusion process to improve feature selection, and the present application has the advantages of fast detection speed, high detection precision, strong noise robustness and good scene adaptability, and has wide application scenarios in high-quality strip steel production scenes, strip steel surface defect detection, micro target detection and other fields.

Description

Technical Field

[0001] This invention relates to the field of surface defect detection technology, specifically to a system and method for detecting minute defects on the surface of steel strips. Background Technology

[0002] As a fundamental product of the steel industry, strip steel is widely used in construction, automobile manufacturing, and daily consumer goods. Surface defects in strip steel can lead to quality problems in work-in-process and downstream products. Typically, image-based automated vision inspection (AVI) instruments are used to expose and locate these defects. Among various types of surface defects, micro-defects are considered one of the most serious threats to steel surface quality due to their typical periodicity and high concurrency. If not inspected promptly, these defects can be hidden due to factors such as process deviations, and will worsen over time, leading to more serious risks. AVI systems can detect and record this deterioration process, thereby controlling strip steel quality problems at an early stage.

[0003] However, in actual production lines, due to the high temperatures (approximately 600℃-800℃), imaging equipment must be installed at a considerable distance (approximately 2 meters) from the strip surface. The imaging size of minute defects is always extremely small, occupying less than 1% of the entire original image area. These objects are not only small in size but also lack characteristic information regarding color or texture. The inherent problems of small defect size and low contrast make defect detection tasks even more challenging. Therefore, this basic but “difficult” detection of ultra-small targets limits the ability to improve the final AVI performance in this specific industrial scenario. Summary of the Invention

[0004] The present invention addresses the challenge that existing technologies struggle to accurately and efficiently identify minute surface defects with low contrast in industrial settings, and provides a system and method for detecting minute defects on the surface of steel strips.

[0005] The objective of this invention can be achieved through the following technical solutions: On the one hand, this application provides a strip steel surface micro-defect detection system, which includes acquiring original images from the strip steel production line, creating a hot-rolled strip steel micro-defect detection database, a backbone feature extraction network, an orthogonal context attention module, a cross-adaptive aggregation module, and a decoupled detection head.

[0006] The backbone feature extraction network is configured to extract features from the image, transforming the original input image into a multi-layer feature map;

[0007] The orthogonal context attention module is configured to simultaneously enhance contextual detail information from both vertical and horizontal directions, thereby enhancing the feature representation of small targets.

[0008] The cross-adaptive aggregation module is configured to use cross-semantic priors to guide the selection of foreground features and mitigate the impact of amplified background noise.

[0009] The decoupled detection head is configured to generate defect categories and location information.

[0010] On the other hand, this application provides an implementation method for a strip steel surface micro-defect detection system, including the following steps:

[0011] S10: Obtain raw images from the hot-rolled strip production line and create a database for detecting minor defects in hot-rolled strip.

[0012] S20: Divide the hot-rolled strip steel micro-defect detection database into training and validation sets, and input them into the network for training;

[0013] S30: Construct a strip steel surface micro-defect detection system. The strip steel surface micro-defect detection system model consists of a backbone feature extraction network, a neck feature fusion network, and a head network. The backbone feature extraction network consists of CSPDarknet53. The neck feature fusion network consists of an orthogonal context attention module and a cross-adaptive aggregation module. The head network consists of a decoupled detection head with decoupled regression and classification branches.

[0014] S40: The pre-trained backbone feature extraction network extracts features from images in the hot-rolled strip steel micro-defect detection database;

[0015] S50: Generate shallow feature maps from the backbone feature extraction network in step S40. C 3 and C 4. Input the orthogonal context attention module to generate an improved feature map with enhanced target context information. and ;

[0016] S60: Generate a deep feature map from the backbone feature extraction network in step S40. C 5 and the improved feature map generated in step S50 and The input cross-adaptive aggregation module utilizes cross-semantic priors to guide the selection of foreground features and mitigates the impact of amplified background noise, generating an adaptively summarized feature map for identifying minor defects.

[0017] S70: The feature map generated in step S60, which adaptively summarizes the identification of minor defects, is input into the decoupled detection head to generate a configuration of defect category and location information, thereby detecting minor defects in hot-rolled strip steel.

[0018] As a further improvement to the above scheme,

[0019] Preferably, the orthogonal context attention module in step S50 is composed of three convolutions with different dilation rates concatenated together, and simultaneously inputs a 1×k horizontal variable convolution and a k×1 vertical variable convolution. After being processed by the Sigmoid function, the convolutions are added element by element and then attention weighted.

[0020] Preferably, the specific process for generating the improved feature map is as follows:

[0021]

[0022] in, C s This represents the enhanced contextual information obtained by sequentially applying convolutions with different dilation rates and orthogonal attention.

[0023] Preferably, the enhanced feature map C s By analyzing cascade features C cat Application of orthogonal attention score ( α row , α col To generate, specifically:

[0024]

[0025] in, ψ (·) represents the operations of sequential partitioning and element-wise summation, and the orthogonal attention score. α row , α col ) represent the weights of feature attention in the horizontal and vertical directions, respectively; specifically:

[0026]

[0027] σ (·) represents the Sigmoid function. W r This represents a 1×k deformable convolution in the horizontal direction. W c This represents a k×1 deformable convolution in the vertical direction.

[0028] Preferably, deformable convolution DCNv3 is introduced to model attention weights within a specific range in the horizontal or vertical direction:

[0029]

[0030] in, G It is the number of groups. w g Represents the position-independent projection weights in each group. mgk It is the normalized modulation factor of the k-th sampling point in the g-th group. x g The input feature map represents a slice. p 0 and p k Representing the current and k-th sampling points respectively, Δ p gk It is the offset corresponding to the grid sampling position pk in the g-th group.

[0031] Preferably, the cross-adaptive aggregation module in step S60 consists of a set of two pointwise convolutions, a sigmoid function operation, a 1×1 convolution operation, a bilinear upsampling interpolation operation, global average pooling, two 1×1 convolutional layer sequences, batch normalization, and a ReLU activation function.

[0032] Preferably, the cross-adaptive aggregation module adaptively summarizes the minor defect identification features:

[0033]

[0034] in, M i This represents a 1×1 convolution operation. Indicates element-wise multiplication. This indicates summation of elements.

[0035] Preferably, by Φ i (·) Extract detailed information from low-level features to obtain the detail-guided attention vector. Z i :

[0036]

[0037] in, Φ i (·) represents a set of two pointwise convolutions and a sigmoid function operation, W1∈R. C×C / r The weights associated with the first pointwise convolution, W2∈R C / r×C With the weights of the second pointwise convolution, the function σ i This represents the Sigmoid operation;

[0038] Preferably, the channel projection function Θ (·) Applied to deep feature maps Extract semantic information from deep feature maps to obtain semantically guided attention vectors. Y i+1 :

[0039]

[0040] In the formula Θ (·) denotes the channel projection operation, W1∈R C×C / r W2 represents the learnable weights of the first convolutional layer. C / r×C This represents the learnable weights of the second convolutional layer, with a reduction ratio of 4. P avg (·) δ (·)and σ (·) represent average pooling, ReLU activation function, and Sigmoid function, respectively. It uses convolution and upsampling operations to process feature maps. This yields a channel with size C and a size of W×H. Feature map, as shown below:

[0041]

[0042] in, X (·) indicates that in 1×1 convolution operation is applied continuously on the feature map W (·) and bilinear upsampling interpolation operations U (·)

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] The strip surface micro-defect detection system and method provided by this invention can better solve the problem of micro-defect detection under low-contrast images. Inspired by the prior information that the horizontal distribution of micro-defects is more uneven than the vertical distribution, an orthogonal contextual attention module (OCA) is carefully designed between the backbone and neck network to better capture micro-target features by enhancing contextual representation. In addition, a cross-adaptive aggregation module (CAA) is also constructed to strengthen the selection of discriminative features by using cross-layer semantic priors during feature fusion.

[0045] Compared with existing defect detection algorithms, the method provided by this invention can accurately detect tiny defects on the surface of steel strips, solving the problem of tiny defects being small in size and having low contrast in specific industrial scenarios. This effectively improves the performance bottleneck of automatic visual inspection (AVI) instruments in detecting ultra-small targets under low contrast conditions. Attached Figure Description

[0046] Figure 1 A flowchart of the strip surface micro-defect detection system and method provided by the present invention;

[0047] Figure 2 This is a schematic diagram of the overall model structure of the strip steel surface micro-defect detection system and method provided by the present invention;

[0048] Figure 3 A schematic diagram of the orthogonal context attention module of the strip surface micro-defect detection system and method provided by the present invention;

[0049] Figure 4 This is a schematic diagram of the cross-adaptive aggregation module of the strip surface micro-defect detection system and method provided by the present invention. Detailed Implementation

[0050] Please refer to the following: Figures 1 to 4 This invention provides a system and method for detecting minute defects on the surface of steel strip, specifically including an orthogonal contextual attention module and a cross-adaptive aggregation module.

[0051] Tiny defects contain almost no intrinsic information such as texture, shape, and color. This means that relying solely on target-centric feature representations is insufficient, especially for tiny targets. Therefore, extracting discriminative features for these objects is crucial. In reality, targets often coexist with their surrounding scene or other targets; accurately modeling this contextual relationship can significantly improve the detection of tiny targets. Unlike other scenarios, due to the influence of the production environment and processes, strip surface images exhibit vertical similarity and horizontal differences around tiny defects. Given this inherent production prior, this invention proposes an Orthogonal Contextual Attention (OCA) module, which simultaneously enhances contextual detail information in both vertical and horizontal directions, thereby strengthening the feature representation of small targets.

[0052] Orthogonal context attention modules inevitably lead to over-enhancement of background regions. Effective small object detection requires selecting discriminative information from redundant features during feature fusion. Commonly used feature fusion networks, such as feature pyramid networks and path aggregation networks, typically fuse features from adjacent layers through linear addition or concatenation. However, since over-enhancing background information is unreliable, linear aggregation makes this information easily propagated, affecting feature selection and interfering with object detection, especially for small objects. Therefore, this invention proposes a cross-adaptive aggregation module (CAA), which utilizes cross-semantic priors to guide foreground feature selection and mitigate the impact of background noise amplification. This module takes low-level details and high-level semantic features as input and generates detail-guided attention and semantic-guided attention, adaptively summarizing features for discriminative small defects.

[0053] Figure 1The algorithm flow is given. First, the original images are obtained from the strip steel production line to create a database for detecting minor defects in hot-rolled strip steel. Then, the images are input into the backbone feature extraction network for feature extraction. Next, the orthogonal context attention module and the cross-adaptive aggregation module are used respectively. Finally, the defect category and location information configuration are generated by decoupling the detection head, thereby detecting minor defects in hot-rolled strip steel.

[0054] The method provided by this invention can accurately detect minute defects in strip steel while ensuring real-time detection.

[0055] Specifically, the implementation method of the strip steel surface micro-defect detection system proposed in this invention includes the following steps:

[0056] S10: Obtain raw images from the hot-rolled strip production line and create a database for detecting minor defects in hot-rolled strip.

[0057] S20: Divide the hot-rolled strip steel micro-defect detection database into training and validation sets, and input them into the network for training;

[0058] S30: Construct a strip steel surface micro-defect detection system. The strip steel surface micro-defect detection system model consists of a backbone feature extraction network, a neck feature fusion network, and a head network. The backbone feature extraction network consists of CSPDarknet53. The neck feature fusion network consists of an orthogonal context attention module and a cross-adaptive aggregation module. The head network consists of a decoupled detection head with decoupled regression and classification branches.

[0059] S40: The pre-trained backbone feature extraction network extracts features from images in the hot-rolled strip steel micro-defect detection database. CSPDarknet53 is the backbone feature extraction network in the strip steel surface micro-defect detection system model. (Refer to...) Figure 2 The input image is first fed into a Focus network for feature extraction, and then into a Conv2D-BN-SiLU module consisting of 1×1 and 3×3 convolutions. Subsequently, it is fed into four Resblock bodies to generate four valid feature maps. The CspLayer network structure consists of a backbone and residual edge parts. The backbone is composed of basic residual convolutions, while the residual edge parts undergo minimal processing before being combined with the output. When the input is 640×640, the scales of the four feature maps are 160×160, 80×80, 40×40, and 20×20, respectively. Three feature maps located in the middle layer (80×80), lower middle layer (40×40), and bottom layer (20×20) are used to construct the subsequent feature fusion network. Notably, an SPP module is used in the deep feature maps, expanding the network's receptive field by paralleling max pooling of different sizes to improve feature extraction capabilities.

[0060] S50: Generate a shallow feature map from the backbone feature extraction network in step S40. C 3 and C 4. Input the orthogonal context attention module to generate an improved feature map with enhanced target contextual information. and ;refer to Figure 2 The orthogonal context attention module consists of three convolutions with different dilation rates concatenated, simultaneously inputting a 1×k horizontally variable convolution and a k×1 vertically variable convolution. After being processed by the sigmoid function, the elements are summed element-wise before attention weighting. This structure enhances contextual details in both the vertical and horizontal directions, thereby improving the feature representation of small objects.

[0061] Specifically, in step S50 Figure 3 The structure of the Orthogonal Context Attention (OCA) module is shown, where HDC represents Horizontal Deformable Convolution and VDC represents Vertical Deformable Convolution. Convolutions with different dilation rates are used to collect contextual information from multiple receptive fields. Subsequently, orthogonal attention is used to enhance the target contextual information both vertically and horizontally. The OCA module is located after the backbone shallow feature map, which contains more subtle target features, such as the shallow feature map itself. C 3 and C 4. Specifically, the orthogonal context attention module takes as input a shallow feature map containing detailed information about tiny objects. C i To generate improved feature maps with enhanced target contextual information. This process is represented as follows:

[0062] in, C s This represents the enhanced contextual information obtained by sequentially applying convolutions with different dilation rates and orthogonal attention. This process helps to highlight the contextual details of small objects.

[0063] More precisely, through the cascading features C cat Application of orthogonal attention score ( α row , α col To generate enhanced feature maps C s First, cascade features. C cat The output is obtained by aggregating dilated convolutions with different dilation rates, where rate = 1, 2, 3. Then, the horizontal attention score is obtained by Hadamard product. α rowand vertical attention score α col Applied to C cat Finally, the results are split and merged. C s As shown in the formula:

[0064] in, ψ (·) indicates the operations of sequential partitioning and element-wise summation. Orthogonal attention score ( α row , α col ) represent the weights of the feature attention in the horizontal and vertical directions, respectively. Their calculation formulas are as follows:

[0065] here, σ (·) represents the Sigmoid function. W r This represents a 1×k deformable convolution in the horizontal direction. W c This represents a k×1 deformable convolution in the vertical direction. Deformable convolution DCNv3 is introduced to model attention weights within a specific range in the horizontal or vertical direction. Simply put, deformable convolution adds an offset to the traditional convolution operation. This allows the convolution kernel to dynamically adjust according to the shape and position of the input features, thereby extracting more accurate and complete relevant object features. Deformable convolution DCNv3 enhances deformable convolution by sharing weights among convolutional neurons, introducing multiple sets of mechanisms, and normalizing the modulation scalar during sampling. This helps to learn and model input features more accurately, resulting in more accurate results. The expression for deformable convolution DCNv3 can be represented as:

[0066] in, G It is the number of groups. w g Represents the position-independent projection weights in each group. m gk It is the normalized modulation factor of the k-th sampling point in the g-th group, which is normalized along the K-th dimension using the Softmax function. x g The input feature map represents a slice. p 0 and p k Representing the current and k-th sampling points respectively, Δ p gk It is the offset corresponding to the grid sampling position pk in the g-th group.

[0067] S60: Generate a deep feature map from the backbone feature extraction network in step S40. C Improved feature maps generated in S5 and S50 and The input cross-adaptive aggregation module utilizes cross-semantic priors to guide the selection of foreground features and mitigates the impact of amplified background noise, generating an adaptively summarized feature map for identifying minor defects; see reference... Figure 4 The cross-adaptive aggregation module consists of a set of two pointwise convolutions, a sigmoid function operation, a 1×1 convolution operation, a bilinear upsampling interpolation operation, global average pooling, two 1×1 convolutional layer sequences, batch normalization, and a ReLU activation function. This structure takes low-level details and high-level semantic features as input and generates detail-guided attention and semantic-guided attention, adaptively summarizing features for identifying minor defects.

[0068] Specifically, in step S60 Figure 4 This demonstrates the structure of the Cross-Adaptive Aggregation Module (CAA), a set of [ , ] is defined as the set of encoded inputs. Among them, Represents low-level characteristics. Representing high-level features (i=3,4). To obtain detail-guided attention vectors, in the shallow feature map... Each spatial location point in the vector is used to summarize and encode channel features. Detail-guided attention vectors are then used. Z i The calculation formula is as follows:

[0069] in, Φ i (·) represents a set of two pointwise convolutions and a sigmoid function operation. The weights associated with the first pointwise convolution are denoted as W1∈R. C×C / r The weights of the second pointwise convolution are represented as W2∈R C / r×C .function σ i This represents the Sigmoid operation. The output detail attention vector is used. Z i Weighted deep feature map It identifies and distinguishes information by providing rich details.

[0070] deep feature map The channel dimension in the input deep feature map is encoded as a separate object response, serving as a semantically guided attention vector. The input deep feature map has 2C channels and a size of W / 2 × H / 2. To align the feature representations, It uses convolution and upsampling operations to process deep feature maps. This yields a channel with size C and a size of W×H. Feature map, as shown below:

[0071] in, X (·) indicates that in 1×1 convolution operation is applied continuously on the feature map W (·) and bilinear upsampling interpolation operations U (·) to ensure with feature Figure 1 To achieve the desired channel and spatial resolution, the channel projection function is then applied. Θ (·) Applied to deep feature maps This yields a semantically guided attention vector. Y i+1 Its definition is as follows:

[0072] In the formula Θ (·) denotes the channel projection operation, which includes global average pooling, two 1×1 convolutional layer sequences, batch normalization, and the ReLU activation function. W1∈R C×C / r W2 represents the learnable weights of the first convolutional layer. C / r×C This represents the learnable weights of the second convolutional layer. The reduction rate is set to 4. Furthermore, P avg (·) δ (·)and σ (·) represent average pooling, ReLU activation function, and Sigmoid function, respectively. Then, in the shallow feature map... Using semantically guided attention vectors Y i+1 This allows it to acquire more semantic information, reduces the influence of unreliable features, and facilitates the selection of shallow foreground information.

[0073] Finally, CAA adaptively summarizes the features for identifying minute defects, and the output results are as follows:

[0074] in, M i This represents a 1×1 convolution operation, used for non-linear processing of fused features. Indicates element-wise multiplication. This indicates summation of elements.

[0075] S70: Input the feature map generated in step S60 into the decoupled detection head to generate defect category and location information configuration, thereby detecting minor defects in hot-rolled strip steel.

[0076] As an experimental example of this invention, the Steel-Tiny database test set, a database for detecting small roll marks on hot-rolled strip steel from Central South University, was used for testing. Several other popular target detection methods, such as Faster-RCNN, Cascade RCNN, YOLOv5, YOLOv6, YOLOv7, and YOLOv8, were also introduced as comparison groups to highlight the superiority of this method. Table 1 summarizes the quantitative evaluation results of different methods. The strip steel surface small defect detection system provided by this invention achieved an average accuracy (mAP) of 71.7%, surpassing other state-of-the-art (SOTA) target detection methods in terms of detection accuracy. Regarding time complexity, the strip steel surface small defect detection system provided by this invention achieves a speed of 77.3 frames per second (FPS). Although it may be slightly slower than some models in the YOLO series, it still meets the real-time detection requirements of strip steel production lines. In summary, the experimental results demonstrate the excellent performance of the strip steel surface small defect detection system in detecting low-contrast small roll marks, meeting the requirements of practical applications.

[0077] Table 1 Quantitative evaluation results of different methods

[0078]

[0079] The present invention provides a network and method for detecting minute defects on the surface of strip steel, which enables accurate and efficient identification of minute defects on the surface of low-contrast strip steel in industrial settings.

[0080] First, this invention proposes an orthogonal context attention module. Based on the inherent prior of production, namely the vertical homomorphism and horizontal difference of small defects, the orthogonal context attention module enhances the contextual detail information from both vertical and horizontal directions, thereby enhancing the feature representation of small targets.

[0081] Furthermore, this invention proposes a Cross-Adaptive Aggregation (CAA) module, which utilizes cross-semantic priors to guide the selection of foreground features and mitigate the impact of amplified background noise. This module takes low-level details and high-level semantic features as input and generates detail-guided attention and semantic-guided attention, adaptively summarizing features for identifying minor defects.

[0082] Compared with several popular target detection algorithms, this invention can detect minute defects on the surface of strip steel more accurately while ensuring real-time detection.

[0083] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A system for detecting minute defects on the surface of strip steel, characterized in that, This includes acquiring raw images from the strip steel production line to create a database for detecting minute defects in hot-rolled strip steel, a backbone feature extraction network, an orthogonal context attention module, a cross-adaptive aggregation module, and a decoupled detection head; The backbone feature extraction network is configured to extract features from the image, transforming the original input image into a multi-layer feature map; The orthogonal context attention module is configured to simultaneously enhance contextual detail information from both vertical and horizontal directions, thereby enhancing the feature representation of small targets. The orthogonal context attention module is composed of three convolutions with different dilation rates concatenated together, and simultaneously inputs a 1×k horizontal variable convolution and a k×1 vertical variable convolution. After being processed by the Sigmoid function, the convolutions are added element by element and then attention weighting is performed. The cross-adaptive aggregation module is configured to use cross-semantic priors to guide the selection of foreground features and mitigate the impact of amplified background noise. The cross-adaptive aggregation module consists of a set of two pointwise convolutions, a sigmoid function operation, a 1×1 convolution operation, a bilinear upsampling interpolation operation, a global average pooling, two 1×1 convolutional layer sequences, batch normalization, and a ReLU activation function. The cross-adaptive aggregation module adaptively summarizes the features for identifying minor defects: in, M i This represents a 1×1 convolution operation. Indicates element-wise multiplication. This indicates summation of elements; Will Φ i (·) Applied to shallow feature maps Extract detailed information from the shallow feature map to obtain the detail-guided attention vector. Z i : in, Φ i (·) represents a set of two pointwise convolutions and a sigmoid function operation, W1∈R. C×C / r The weights associated with the first pointwise convolution, W2∈R C / r×C With the weights of the second pointwise convolution, the function σ i This represents the sigmoid operation; Channel projection function Θ (·) Applied to deep feature maps Extract semantic information from deep feature maps to obtain semantically guided attention vectors. Y i+1 : In the formula Θ (·) denotes the channel projection operation, W1∈R C×C / r W2 represents the learnable weights of the first convolutional layer. C / r×C This represents the learnable weights of the second convolutional layer, with a reduction ratio of 4. P avg (·) δ (·)and σ (·) represent average pooling, ReLU activation function, and Sigmoid function, respectively. It uses convolution and upsampling operations to process feature maps. This yields a channel with size C and a size of W×H. Feature map, as shown below: in, X (·) indicates that in 1×1 convolution operation is applied continuously on the feature map W (·) and bilinear upsampling interpolation operations U (·); The decoupled detection head is configured to generate defect categories and location information.

2. A method for implementing a strip steel surface micro-defect detection system, based on the method for implementing the strip steel surface micro-defect detection system according to claim 1, characterized in that, Includes the following steps: S10: Obtain raw images from the hot-rolled strip production line and create a database for detecting minor defects in hot-rolled strip. S20: Divide the hot-rolled strip steel micro-defect detection database into training and validation sets, and input them into the network for training; S30: Construct a strip steel surface micro-defect detection system. The strip steel surface micro-defect detection system model consists of a backbone feature extraction network, a neck feature fusion network, and a head network. The backbone feature extraction network consists of CSPDarknet53. The neck feature fusion network consists of an orthogonal context attention module and a cross-adaptive aggregation module. The head network consists of a decoupled detection head with decoupled regression and classification branches. S40: The pre-trained backbone feature extraction network extracts features from images in the hot-rolled strip steel micro-defect detection database; S50: Generate shallow feature maps from the backbone feature extraction network in step S40. C 3 and C 4. Input the orthogonal context attention module to generate an improved feature map with enhanced target context information. and ; S60: Generate a deep feature map from the backbone feature extraction network in step S40. C 5 and the improved feature map generated in step S50 and The input cross-adaptive aggregation module utilizes cross-semantic priors to guide the selection of foreground features and mitigates the impact of amplified background noise, generating an adaptively summarized feature map for identifying minor defects. S70: The feature map generated in step S60, which adaptively summarizes the identification of minor defects, is input into the decoupled detection head to generate a configuration of defect category and location information, thereby detecting minor defects in hot-rolled strip steel.

3. The implementation method of the strip steel surface micro-defect detection system according to claim 2, characterized in that, The specific process of generating the improved feature map is as follows: in, C s This represents the enhanced contextual information obtained by sequentially applying convolutions with different dilation rates and orthogonal attention.

4. The implementation method of the strip steel surface micro-defect detection system according to claim 3, characterized in that, Enhanced feature maps C s By analyzing cascade features C cat Application of orthogonal attention score ( α row , α col To generate, specifically: in, ψ (·) represents the operations of sequential partitioning and element-wise summation, and the orthogonal attention score. α row , α col ) represent the weights of feature attention in the horizontal and vertical directions, respectively; specifically: σ (·) represents the Sigmoid function. W r This represents a 1×k deformable convolution in the horizontal direction. W c This represents a k×1 deformable convolution in the vertical direction.

5. The implementation method of the strip steel surface micro-defect detection system according to claim 4, characterized in that, Introducing deformable convolution DCNv3 to model attention weights within a specific range in the horizontal or vertical directions: in, G It is the number of groups. w g Represents the position-independent projection weights in each group. m gk It is the normalized modulation factor of the k-th sampling point in the g-th group. x g The input feature map represents a slice. p 0 and p k Representing the current and k-th sampling points respectively, Δ p gk It is the offset corresponding to the grid sampling position pk in the g-th group.

Images