Defect lightweight detection model and method based on multi-scale dynamic fusion and application

Abstract

The invention discloses a defect lightweight detection model and method based on multi-scale dynamic fusion and application, and the model comprises the steps: constructing a multi-scale feature enhancement module, improving the receptive field and multi-scale modeling capability through large kernel convolution, multi-scale cavity convolution and structural re-parameterization, and combining with a multi-scale-cavity rate mapping mechanism to obtain a multi-scale feature enhancement module; through predefined cavity convolution configuration, a network receptive field can quantitatively cover defect areas with different physical sizes, and self-adaptive matching of the cavity rate and the defect scale is realized. According to the invention, an adaptive feature fusion module is optimized, context sensing feature fusion is realized again through dynamic multi-path convolution and introduction of a convolution gating mechanism, the calculation overhead is reduced while the precision is maintained, lightweight deployment is realized, and the method is suitable for large-scale popularization and application. The technical problems that in an existing wheel set tread defect detection model, the small-scale defect omission ratio is high, complex working condition false detection is frequent, and light weight and precision are difficult to balance are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of surface defect detection, and more particularly to a defect lightweight detection model and method based on multi-scale dynamic fusion and application. BACKGROUND

[0002] Wheelset as a key bearing component of train, its metal tread state directly affects the stability and safety of train operation. Long-term service tread often appears defects such as pitting and peeling, if not detected in time, may cause vibration, noise and even shaft breakage accident, causing serious safety hazards. The traditional manual inspection and sensor signal based method are difficult to meet the needs of high-density railway transportation in efficiency and accuracy.

[0003] In recent years, deep learning driven visual detection has shown significant advantages in industrial defect recognition. This method can learn and extract multi-level and abstract features from massive image data, making it able to identify subtle and complex defects that are difficult for the human eye to detect or traditional algorithms to describe. It has stronger adaptability to environmental light changes, product position shifts, angle rotations, background disturbances, and more stable detection results. However, existing methods still face the following challenges when applied to surface defect detection: (1) Insufficient multi-scale defect feature modeling, unable to balance small size pitting and large area peeling features; (2) Limited context information capture ability, making it difficult to achieve robust detection in complex backgrounds; (3) There is a contradiction between lightweight design and detection accuracy, as improving accuracy often leads to a decrease in inference speed. Therefore, how to balance detection accuracy, inference speed and model lightweight is still a core problem.

[0004] In CN120259180A, a beer bottle cap production defect visual detection method based on improved YOLOv10n is disclosed. After median filtering processing of the collected data image, a double fuzzy contrast reservation image processing algorithm is designed to better preserve detail information and enhance defect texture information of the image while denoising. C2f-CMAC module and FPSC module are designed to improve the defect feature information capture and fusion ability of the model through CA attention and pyramid feature fusion structure in the module. LSWD-Detect module is designed to replace the original detection head of YOLOv10n network, using shared parameter convolution, group normalization and depth separable convolution to achieve lightweight of the detection head parameters while slightly improving the detection accuracy. At the same time, MPDIoU loss function is introduced in YOLOv10n network to replace the CIoU loss function of the network model, accelerating the convergence speed and model inference speed. This patent method realizes the improvement of model detection accuracy, reduces the miss and false detection rate of products, and realizes the parameter lightweight of the model, improving the portability of the model.

[0005] However, compared with the defect detection scheme for the tread surface of the wheel set, the method still has the following limitations: 1. Insufficient multi-scale feature modeling capability: the CA attention and pyramid feature fusion structure adopted are relatively general, and are not designed for the characteristics of large defect size span (from pixel-level pitting to centimeter-level peeling) of the tread surface of the wheel set, lack the mechanism of self-adaptive adjustment of the receptive field, and it is difficult to simultaneously consider the accurate detection of small targets and large-scale defects.

[0006] 2. Insufficient context information fusion mechanism: under complex background (such as oil stains, rust, uneven illumination), the model lacks dynamic and adaptive context perception and fusion capability, and has limited robustness to edge blur and morphologically variable tread defects.

[0007] 3. The balance point of lightness and precision needs to be optimized: the lightness design mainly focuses on the detection head part, and does not perform global optimization on the core path of feature extraction and fusion. While pursuing parameter reduction, the feature discrimination ability for subtle defects may be sacrificed, and the contradiction between “lightness” and “high precision” in industrial detection cannot be completely solved. SUMMARY

[0008] To overcome the problems such as high miss detection rate of small-scale defects (such as micro pitting), frequent false detection under complex background (oil stains, reflection, oxidation layer), poor receptive field adaptability of multi-scale defects, difficulty in balancing lightness and detection precision, and large calculation overhead during feature fusion in the existing defect detection technology for the tread surface of the wheel set, the application provides a defect lightness detection model based on multi-scale dynamic fusion.

[0009] Another technical problem solved by the application is to provide a defect lightness detection method and application based on multi-scale dynamic fusion based on the model.

[0010] The application is implemented by the following technical solutions:

[0011] A defect lightness detection model based on multi-scale dynamic fusion, comprising: A data collection module including a camera or a video camera, which collects defect images of workpieces under different service states and different working conditions.

[0012] A preprocessing module including an image denoising submodule, a geometric transformation submodule, a gray scale enhancement submodule, and an advanced data enhancement submodule, which performs denoising, geometric transformation, and data enhancement processing on the collected data to construct diversified training samples; The multi-scale feature extraction backbone network is composed of four groups of multi-scale feature enhancement modules with different hole configurations, and the multi-scale feature enhancement module includes a multi-scale hole reparameterization module and a hole pyramid attention module for adaptive feature extraction of defect scales. The multi-scale hole reparameterization module includes an original feature extraction branch and a deep feature extraction branch, which adopts a double-path parallel structure. Path one is an identity mapping branch that retains the original feature information through 1x1 convolution. Path two is a multi-scale feature extraction branch composed of a large kernel convolution and a hole convolution with different hole rates in parallel, which captures local details and context information in different ranges. The hole pyramid attention module is composed of a multi-scale hole convolution, a defect-sensitive activation function, and a channel attention fusion. The input features are first formed into multi-scale features through multi-scale hole convolution, then each branch feature is nonlinearly transformed through a defect-sensitive activation function to highlight the defect area and suppress the background. Finally, the channel attention mechanism is used to assign different channel weights to the feature maps of the three branches, and weighted summation is performed, followed by feature integration and output. The adaptive feature fusion module includes a C2f-DCFN module optimized based on a dynamic mixed convolution module, a GSConvs module, and a VoVGSCSPC module, which perform deep interactive fusion of the multi-scale features extracted above, retain multi-scale defect details, and reduce redundancy. The defect detection output module includes three multi-scale detection head sub-modules, a classification branch, a regression branch, and a loss function calculation layer, which receive the features output by the adaptive feature fusion module, perform classification and bounding box regression, and finally output the class and position of the defect.

[0013] Further, in the adaptive feature fusion module, the C2f-DCFN module is obtained by replacing the Bottleneck module of the original c2f module with the DCFN module, which uses DCFN to capture more comprehensive multi-scale context information, enhance context-dependent modeling, and improve asymmetric feature coupling capability. The GSConvs module replaces the shuffle operation with 1x1conv+LeakyReLU activation function based on GSConv, which learns the defect feature channel weight through training, dynamically enhances the key feature response strength, and improves the feature discrimination of small-scale micro-cracks and other defects. VoVGSCSPC is a single GSBottleneckC module obtained by stacking multiple GSBottleneck based on VoVGSCSP, with the hidden channel number fixed at 50% of the output channel, and GSBottleneckC replaced with DWConv as the shortcut branch.

[0014] An industrial defect lightweight detection method based on multi-scale dynamic fusion, comprising the following steps: S1. Collecting image data and pre-processing the image data to obtain an input picture; S2. Multi-scale feature extraction is performed on the input picture; S21. The input picture is processed by dimension reduction to obtain original features X1 and X2 from the channel dimension; S22. The original feature X2 is extracted by convolution and convolution kernel-dilated convolution to obtain shallow features and deep features F dil ; S23. The shallow features and the deep features F dil are weighted and fused to obtain multi-scale features , and the specific formula is:

[0015] is the output feature, i.e., the multi-scale feature ; F std is the shallow feature, and F dil is the deep feature; represents weighted fusion; S24. The multi-scale features are formed into different receptive fields by multi-scale dilated convolution, and a defect-sensitive activation function is introduced to control the response difference between abnormal areas and background through channel threshold, highlighting the defect features, wherein the defect-sensitive activation function is:

[0016] In the formula, T∈R^(C×1×1) is a learnable channel threshold; α controls the negative response slope (less than 1), β controls the overall scaling, and ReLU( ) represents a rectified linear unit; Finally, the channel attention is used to weight and fuse the output features X3 of each branch feature, and the formula is:

[0017] wherein Y is the output feature, i.e., the fusion output feature X3; represents the weight of the i-th branch; the learnable fusion weight, represents the output feature of each branch after defect activation; S25. The fusion output feature X3 is taken as input to repeat steps S22-S24 to obtain a fusion output feature X4; S26. The original feature X1, the multi-scale feature , the fusion output feature X3, the fusion output feature X4, and the quarter feature are spliced and output to obtain an enhanced multi-scale feature P; S3. The multi-scale feature P is first normalized and adapted, then up-sampled and aligned across scales for splicing and fusion, and finally enhanced through adaptive mixed convolution, and finally a discriminative feature is obtained through lightweight integration ; S31. Channel layer normalization is performed on the multi-scale feature, channel adaptation transformation is performed according to the need for cross-scale splicing, and then the space is aligned and spliced and fused to obtain a cross-scale fusion feature; S32. The cross-scale fusion feature is divided into a main branch and an enhanced branch , and the main branch is extracted through convolution and activation to obtain a feature ; S33. The enhanced branch first generates a dynamic weight, then extracts through a multi-morphology DW convolution branch, and finally outputs a feature , which is represented as: ; wherein Y is an output feature, is a corresponding weight of each group, is a DW convolution output; S34. The features and the features are spliced in the channel and unified in the dimension to obtain a feature , and the features and the features are residual connected to obtain a feature ; S35. The feature is integrated through GSConvs and VoVGSCSPC modules to obtain a discriminative feature of the adaptive detection head ; S4. The discriminative feature output by S3 is processed through a classification branch and a regression branch to obtain a detection result of the wheel tread defect.

[0018] Further, the preprocessing includes image denoising, geometric transformation, gray scale enhancement, and advanced data enhancement.

[0019] Further, in the inference stage, the hole convolution group is equivalent to a convolution through the reparameterization method, and the formula is as follows: ; wherein represents the output of the inference, that is, the deep feature F dil , Align( ) represents the mapping of different dilated convolution weights to a unified kernel size, represent the weight of the i-th branch, represent the mapped weights.

[0020] Further, the receptive field R is calculated according to the formula:

[0021] where R represents the effective receptive field, K represents the convolution kernel size, and d represents the dilated convolution expansion rate Further, the deep feature extraction fuses the outputs of different branch dilated convolutions according to learnable weights during training, and the formula is expressed as: , wherein represents the output in the training stage, represents the weight of the i-th branch; * represents the convolution operation; learnable fusion weight, represents the output feature of each branch.

[0022] Further, the normalization formula is:

[0023] wherein X is the input feature, i.e. the enhanced branch , is the mean of the feature channel dimension, is the channel variance, .

[0024] Further, the dynamic weight is represented as: ; is the input feature, is the global average pooling, is convolution operation, is the three sets of weights.

[0025] Further, the specific steps of step S4 include: S41. input the discriminative feature to the detection head, and the classification branch obtains the defect class prediction probability after feature enhancement and class prediction; S42. input again to the regression branch, and obtain the defect candidate frame after feature enhancement, parameter prediction and coordinate decoding; S43. merge the obtained scale candidate frame, and output the final defect detection result after confidence screening and non-maximum suppression.

[0026] The industrial defect lightweight detection method based on multi-scale dynamic fusion is applied to defect detection of a wheelset tread.

[0027] Compared with the prior art, the beneficial effects of the present application are: 1. The multi-scale feature enhancement module is constructed, the receptive field and multi-scale modeling capability are improved through large kernel convolution, multi-scale hollow convolution and structure reparameterization, the "training-multiple branches, inference-single branch" strategy is adopted by fusing the hollow convolution and structure reparameterization technology, the problem that the fixed receptive field convolution kernel is difficult to adapt to the wheelset defect size span is solved. In addition, a multi-scale-hollow rate mapping mechanism is designed, the network receptive field can quantitatively cover the defect area of different physical sizes through the pre-defined hollow convolution configuration, and adaptive matching of the hollow rate and the defect scale is realized.

[0028] 2. The DCFN module of the present application realizes dynamic multi-path convolution and introduces a convolution gating mechanism, and re-implements context-aware feature fusion.

[0029] 3. The adaptive feature fusion module of the present application optimizes the calculation overhead while maintaining the accuracy, and realizes lightweight deployment. Specifically, the C2F-DCFN module is combined with the GSConv and the C2f-DCFN module, the DCFN is used to more comprehensively capture multi-scale context information, enhance the context-dependent modeling and asymmetric feature coupling capability, the GSConv based on lightweight convolution technology is improved, the shuffle operation is replaced by 1x1conv+Leaky ReLU activation function, the defect feature channel weight is learned through training, the key feature response strength is dynamically enhanced, and the feature discriminability of small-scale microcracks and other defects is improved. In combination with VoVGSCSP, multiple GSBottleneck are stacked and simplified into a single GSBottleneckC module, the hidden channel number is fixed to 50% of the output channel, and the GSBottleneckC uses DWConv instead of ordinary convolution as the shortcut branch, reduces the parameter quantity and floating point operation, and adapts to real-time detection requirements. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is a model overall framework diagram in Example 1; Figure 2 is a multi-scale feature enhancement module structure schematic diagram; Figure 3 is a multi-scale hollow reparameterization schematic diagram; Figure 4 is an empty pyramid defect enhancement attention schematic diagram; Figure 5 is an adaptive context fusion module structure schematic diagram; Figure 6is a RDL-Neck schematic diagram; Figure 7 is a C2F structure comparison diagram; Figure 8 is a defect sample example diagram; Figure 9 is a map50 curve diagram of the training process of Example 3; Figure 10 is a GradCAM++ heat map; Figure 11 is a map50 curve diagram of the training process of Example 4 comparative experiment; Figure 12 is a visualization result analysis diagram of Example 4; Figure 13 is a map50 curve diagram of the training process of Example 5 generalization experiment. DETAILED DESCRIPTION

[0031] In order for those skilled in the art to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below. Obviously, the described embodiments are only part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0032] Example 1 The present embodiment provides a defect lightweight detection model based on multi-scale dynamic fusion, such as Figures 1-7 , comprising: The data collection module includes a 200 million pixel area array industrial camera, which collects defect images of the wheel tread under different service states (new wheel set - in-service wheel set - aged wheel set) and different working conditions (indoor maintenance workshop fixed scene, open track dynamic scene).

[0033] The preprocessing module includes an image denoising submodule, a geometric transformation submodule, a gray scale enhancement submodule, and an advanced data enhancement submodule, which denoises, geometrically transforms, and data enhances the collected data to construct diversified training samples.

[0034] The multi-scale feature extraction backbone network is composed of four groups of multi-scale feature enhancement modules with different hole configurations (p2-p5 feature extraction layers). The module combines the self-designed multi-scale hole reparameterization module and the pyramid attention module on the GELAN reuse architecture to realize adaptive feature extraction of defect scales. Among them: The multi-scale feature enhancement module includes a multi-scale hollow reparameterization module and a hollow pyramid attention module, and is used for adaptive feature extraction of defect scales.

[0035] In addition, four groups of hollow convolution groups are designed in the multi-scale hollow reparameterization module according to the distribution characteristics of the wheel tread defect scales, and a multi-scale-hollow rate mapping mechanism is formed, so that four different multi-scale feature extraction modules (p2-p4 layers) of the multi-scale feature extraction backbone network are obtained.

[0036] The hollow pyramid attention module (DPA) is composed of multi-scale hollow convolution, defect-sensitive activation function (DSA) and channel attention fusion. The input feature map is first passed through three 3×3 hollow convolutions with hollow rates of 1, 2 and 4 in parallel to form multi-scale features; then, each branch feature is subjected to nonlinear transformation by a defect-sensitive activation function (DSA) to highlight the defect area and suppress the background; finally, a channel attention mechanism (using a compression-excitation structure) is used to assign different channel weights to the feature maps of the three branches, and weighted summation is performed, and then a standard 3×3 convolution layer is used for feature integration and output.

[0037] The adaptive feature fusion module includes a C2f-DCFN module optimized based on a dynamic hybrid convolution module (DCFN), a GSConvs module and a VoVGSCSPC module. Among them: The C2f-DCFN module is obtained by replacing the Bottleneck module of the original c2f module with a self-designed DCFN module, which uses DCFN to capture more multi-scale context information, enhance context-dependent modeling and asymmetric feature coupling capability.

[0038] The GSConvs is improved based on the lightweight convolution technology GSConv, replacing the shuffle operation with 1×1conv+Leaky ReLU activation function, learning the defect feature channel weight through training, dynamically enhancing the key feature response strength, and improving the feature discrimination of small-scale micro-cracks and other defects.

[0039] VoVGSCSPC is improved on the basis of VoVGSCSP, simplifies multiple GSBottleneck to a single GSBottleneckC module, hides the number of channels to 50% of the output channel, and replaces the ordinary convolution with DWConv as the shortcut branch of GSBottleneckC, reducing the parameter quantity and floating-point operation, adapting to the real-time detection demand.

[0040] The defect detection output module includes three multi-scale detection head sub-modules, a classification branch, a regression branch, and a loss function calculation layer, receives the three scale feature maps (large scale 8x8, medium scale 16x16, and small scale 32x32, corresponding to peeling, cracking, and pitting, respectively) output by the adaptive feature fusion module, performs classification and bounding box regression, and finally outputs the category and position of the defect. The fused multi-scale feature map is input into the detection, classification and bounding box regression are performed, and finally the category and position of the defect are output. Embodiment 2 The embodiment provides a defect lightweight detection method based on multi-scale dynamic fusion, which comprises the following steps: S1. Collect image data of the tread of the wheel set, and perform geometric transformation and data enhancement on the defect image of the tread of the wheel set, and construct diversified training samples. Specifically, the image is randomly rotated, scaled, cropped, brightness adjusted, and noise added, and adaptive histogram equalization is used to enhance the contrast of the defect area, and image mixing and random occlusion technology is used to improve the robustness of the model to complex background, and finally the preprocessed input picture is obtained. S11. Original image acquisition: a 200 million pixel area array industrial camera (for example, model Hikvision MV-CA020-10GM, equipped with an 8mm fixed focus lens) is used to collect the tread image of the wheel set at a frame rate of not less than 15fps in the scene of the railway maintenance workshop, open track, etc.

[0041] S12. Basic preprocessing: sequentially perform denoising, geometric transformation, and gray scale enhancement on the collected pictures to obtain the basic preprocessing pictures. Wherein: Image denoising: the collected original image is first subjected to 3x3 Gaussian filtering (standard deviation σ=0.8) to eliminate Gaussian noise, and then subjected to 5x5 median filtering to process salt and pepper noise, and the defect edge details are preserved; Geometric transformation: the denoised image is subjected to-15°~15° random rotation (the blank area is filled with the average gray value of the tread), 0.8~1.2 times random scaling (bilinear interpolation is used to maintain the aspect ratio), and 70%~90% area random cropping (ensure that the cropped area contains defects, and the non-defect image is randomly cropped); Gray scale enhancement, using contrast limited adaptive histogram equalization (CLAHE), setting the clipping threshold 2.0, histogram bin number 256, enhancing the gray scale difference between defects and background; S13. Advanced data enhancement: adopt the strategy of combining classic hybrid enhancement with the latest generative enhancement to construct training samples covering multiple defects, multiple disturbances, and multiple scenes: through the Mixup technology, linear interpolation fusion of images and labels is performed by Beta distribution sampling (λ∈[0.1, 0.4]), which enhances the model's understanding of the smoothness of defect feature distribution; with the help of CutMix, different image local areas are randomly cut and pasted with a rectangular mask, and the labels are mixed according to the mask area ratio, while the local structure information of defects is preserved; Mosaic is used to randomly scale and splice four images of different defect types, and through linear transition fusion, the multi-defect superposition scene is simulated. At the same time, a defect generative adversarial network is introduced to generate synthetic images containing crack bifurcation and small point erosion groups using real samples as seeds to supplement rare defect samples; dynamic on-site disturbance simulation is performed, such as random execution of oil pollution coverage, vibration blur, and oxidation layer texture addition. Through the above methods, the diversity and scene adaptability of the training data are effectively improved by expanding the samples.

[0042] S2. Perform multi-scale feature extraction on the input picture obtained by preprocessing in S1, and generate P2 to P5 feature maps (downsampling by 4, 8, 16, and 32 times, respectively) through four multi-scale feature enhancement modules in series. The core process of each multi-scale feature enhancement module is as follows: S21. After reducing the dimension of the input picture feature through 1×1 convolution, divide it into original feature X1 and original feature X2 from the channel dimension.

[0043] S22. Preliminarily extract multi-scale features from X2 through MDR. MDR includes two parallel paths, which generate multi-scale representations through "shallow feature preservation + deep multi-scale capture". The specific process is as follows: S221. Extract shallow features from X2 through 1×1 convolution , which is used to preserve the original information of X2 and avoid the loss of details caused by deep extraction.

[0044] S222. Capture more rich deep features F dil through parallel branches composed of large convolution kernels and multi-scale dilated convolution. During training, the outputs of different branch dilated convolutions are weighted and fused according to the learnable weights, and the formula is expressed as: , wherein: output of training stage, weight of the i-th branch; * represents convolution operation; learnable fusion weight, output of each branch. In the inference stage, the group of dilated convolutions is equivalent to a convolution by reparameterization method to reduce model latency and complexity, and the deep feature F is obtained dil , the formula is as follows: ; wherein: output of inference, i.e. deep feature F dil , Align( ) represents mapping different dilated convolution weights to a unified convolution kernel size, weight of the i-th branch, mapped weight.

[0045] According to the multi-scale dilated physical mapping table:

[0046] The MDRs of the four DRADPELANs in S2 respectively adopt large kernel convolutions of 13, 11, 9, and 7 corresponding to the dilated convolution groups, which cover the effective receptive fields of 5-11, 5-11, 5-7, and 3-5 through differentiated convolution kernels and dilated rates. The effective receptive field R is calculated according to the formula:

[0047] wherein: R represents the effective receptive field, K represents the convolution kernel university, and d represents the dilated convolution inflation rate.

[0048] S223. The shallow feature is weighted and fused with the deep feature to obtain the multi-scale feature , and the specific formula is:

[0049] output feature, i.e. multi-scale feature ; F std is the shallow feature, and F dil is the deep feature; weighted fusion S23. The multi-scale feature in S223 is subjected to DAP for multi-scale feature enhancement; S231. The multi-scale feature is respectively subjected to 3 dilated convolution branches (with expansion rates of 1, 2, and 4) to form different receptive field coverage from small to large scale defects; S232. Introducing a custom defect-sensitive activation function to regulate the response difference between abnormal areas and the background through channel threshold, highlighting the defect features, where the defect-sensitive activation function is:

[0050] where T ∈ R^(C×1×1) is a learnable channel threshold; α controls the negative response slope (less than 1), β controls the overall scaling, ReLU( ) represents the rectified linear unit; S233. Using channel attention to weight and fuse the output features X3 of each branch, which is expressed as:

[0051] where Y is the output feature, i.e., the fused output feature X3; represents the weight of the i-th branch; learnable fusion weight, represents the output feature of each branch after defect activation.

[0052] S24. Taking the fused output feature X3 as input, repeating the steps S22 and S23 to obtain the fused output feature X4; S25. Concatenating the original feature X1, the original feature X2, the fused output feature X3, and the fused output feature X4 to output the multi-scale feature enhanced feature map P2; S26. Generating feature maps P3-P5 using a similar process as P2, with the difference being the different configuration of the MDR hollow convolution group.

[0053] S3. Processing P3 (80×80×256, down-sampled by 8 times), P4 (40×40×512, down-sampled by 16 times), and P5 (20×20×1024, down-sampled by 32 times) features output by S2 through the full-link process of “normalization adaptation → cross-scale up-sampling concatenation → C2f-DCFN enhancement interaction → lightweight integration”, to preserve multi-scale defect details and reduce redundancy, outputting large, medium, and small three scales of high discriminative features for detection 、 、 ; S31. Normalizing and preliminarily transforming the multi-scale features; For the feature maps P2-P5 output by S2, first eliminate the distribution difference and then adapt to the subsequent splicing needs, the steps are as follows: Layer normalization: Perform normalization on the channel dimension of each feature map separately, the formula is:

[0054] where X denotes input features, Y denotes output features, is the mean of the feature channel dimension, is the channel variance, (avoiding denominator approaching 0); by normalization, eliminate the gray scale distribution difference of different scale features (such as the brightness deviation of P3 shallow detail and P5 deep global feature), ensure the balance of feature weight in subsequent fusion.

[0055] Channel adaptation transformation: according to the channel matching requirement of subsequent cross-scale splicing, adjust the feature channel number: P5 (20x20x1024): 1x1 convolution (step 1, padding 0) dimension reduction to 512 channels, get (20x20x512), matching the channel number (512) of P4; P4 (40x40x512) gets (40x40x256) matching the channel number of P3, laying the foundation for cross-scale splicing.

[0056] S32. Cross-scale up-sampling alignment and splicing fusion adopts the logic of “first up-sampling alignment spatial size → then splicing fusion”, based on the channel adaptation result of S31, realizes complete matching of multi-scale features: S321. Cross-scale fusion of P5 and P4 (up-sampling + splicing): up-sampling alignment spatial size: to (20x20x512) perform nearest interpolation up-sampling 2 times — only change spatial size (20x20→40x40), channel number still keep 512 (consistent with P4), get (40x40x512); at this time and P4 (40x40x512) meet the “spatial size 40x40 + channel number 512” double matching; channel splicing: splice and P4 in channel dimension, get fusion feature (40x40x1024), integrate the global structure information of P5 and the medium-scale defect details of P4.

[0057] S322. splice after C2f-DCFN enhanced interaction, then channel splice with P3 feature in the same step of S321, integrate small-scale defect details to get feature (80x80x512).

[0058] S33. splice cross-scale fusion features and The C2f-DCFN module is input respectively, and the defect features are respectively strengthened by "branch segmentation → adaptive mixed convolution → residual fusion" ( ) and ( ), for example: processing: S331. The is equally divided into 2 branches (each 256 channels): main branch (80x80x256): 3x3 convolution + BN + SiLU activation to obtain (80x80x256), extracting global structure features (such as the overall profile of the wheel tread); enhanced branch (80x80x256): enter the adaptive mixed convolution submodule (DCFN), strengthen the local details of the defect, and DCFN is the core module of adaptive context feature fusion.

[0059] S332. For the enhanced branch (80x80x256), difference defect feature capture is realized by "dynamic weight generation → multi-form DW convolution extraction → weight fusion", and the steps are as follows: S333. Generate dynamic weights: ; Wherein: is the input feature, i.e. the enhanced branch , is the global average pooling, is the convolution operation, is the three groups of weights; S334. Multi-form DW convolution branch extraction

[0060] Where f1, f2, f3 represent 3x3, 1xn, nx1 DWConv convolution respectively; S335. Weighted fusion output to obtain (80x80x256): ; Wherein, Y is the output feature, i.e. , is the corresponding weight of each group, is the DW convolution output.

[0061] S336. Residual interactive fusion to obtain Firstly, the (80x80x256) and (80x80x256) are channel spliced and dimensioned to obtain : Where concat is a channel splicing operation, is a dimension reduction operation that reduces the number of channels from 512 to 256. Then, the (80x80x256) and in S331 are connected by a residual connection, and the two features of the same dimension are added pixel by pixel to finally output the feature ( ), which is consistent with P3 only, and the formula is as follows:

[0062] Where: Add() represents a residual connection.

[0063] S337. Repeat S331 and S336 to obtain ( ); S34. The above obtained ( ), ( ) and P5 (20x20x1024) are integrated by GSConvs and VoVGSCSPC modules to obtain three-level features of the adaptive detection head , , ; S341. The ( ) is directly used as a small-scale defect detection feature ( ); S342. The ( ) is channel spliced with ( ) after GSConvs down-sampling to obtain , and then the middle-scale defect detection feature ( ) is obtained by VoVGSCSPC module; S343. The ( ) and P5 (20x20x1024) are operated after S342 to obtain a large-scale defect detection feature​​​ S4. The third-level features output by S3 are combined by the classification branch and the regression branch of the detection head to obtain the detection result of the wheel tread defect. The detection result of the wheel tread defect is obtained after the classification branch and the regression branch of the detection head are combined and processed.

[0064] S41. The feature map (80x80x256) is input into the classification branch and the regression branch of the detection head, the classification branch predicts the small-scale defect category probability, and the regression branch predicts the bounding box coordinates and confidence, and the small-scale defect candidate box set is obtained after coordinate decoding.

[0065] S411. The feature map (80x80x256) is input into the regression branch, and the defect category prediction probability (score) is obtained after feature enhancement and category prediction.

[0066] S4111. The feature map (80x80x256) is enhanced by two-layer convolution to enhance the category feature:

[0067]

[0068] S4112. The feature map (80x80x256) is input into the regression branch to predict the category probability:

[0069] S412. The feature map (80x80x256) is input into the regression branch, and the small-scale defect candidate box is obtained after feature enhancement, parameter prediction and coordinate decoding, which corresponds to the prediction result in S412.

[0070] S4121. Feature enhancement:

[0071]

[0072] S4122. Bounding box parameter prediction:​​​​​​​​​​​​​​​​​​​​​​​​

[0073] wherein: is a prediction parameter; is a center point relative grid offset; is a width-height logarithmic offset; is a target confidence (0~1).

[0074] S4123. Coordinate decoding Let the feature map grid coordinates be , the down-sampling multiple s = 8, and the reference width and height be , ,

[0075] wherein, is a defect frame center point coordinate; is a defect frame width and height. S413. Combine the results of S411 and S412 to obtain a small-scale defect candidate frame set; S42. After the same operation of S41, a medium-scale defect candidate frame set is obtained; , After the same operation of S41, a medium-scale defect candidate frame set is obtained; F4 (40x40x512) is executed in the same operation of S41, the classification branch predicts the medium-scale defect category probability, the regression branch predicts the boundary frame coordinates and the confidence, and after coordinate decoding, a medium-scale defect candidate frame set is obtained. Similarly, using a large-scale defect candidate frame set is obtained.

[0076] S43. Merge the three candidate frame sets obtained by S41 and S42, and output the final defect detection result (including position, category and comprehensive score) after confidence screening and non-maximum suppression.

[0077] Example 3 This embodiment provides a method for evaluating the performance of a train wheel tread defect detection model, which uses a self-built train wheel tread defect dataset containing multiple types of defects. The dataset contains multiple defect types (such as cracks, wear, pitting, etc.), and is collected under different environmental conditions. For example, Figure 8, the image resolution of the dataset is 600x800 pixels, ensuring that the details of defects in each image can be effectively captured and identified. The dataset contains a total of 927 images, divided into a training set (649 images) and a validation set (278 images) in a 7:3 ratio, with a total of 2574 defects, including 786 peeling defects and 1761 pitting defects. All images are labeled with defect categories and locations; and contain samples of different lighting and shooting angles to ensure that the model adapts to complex actual working conditions and provides a diversified performance verification basis for subsequent experiments.

[0078] To ensure efficient execution of the experiment, the following hardware and software configurations are used, and the training environment is shown in Table 2. To verify the effectiveness of the method proposed in the present application, training and testing were performed on a standard computer system. Through CUDA acceleration, the experiment achieved fast training and inference on a graphics card with a small amount of video memory.

[0079] Table 2

[0080] To improve the efficiency of training and inference, the present application uses CUDA 11.1 for GPU acceleration, significantly reducing the time for model training and prediction. This configuration is suitable for medium-scale deep learning model training, especially when dealing with image datasets, and can to some extent avoid memory overflow problems.

[0081] To comprehensively evaluate the performance of the proposed DRADPELAN-MDPNet model in the wheel tread defect detection task, the present application uses five core indicators widely used in the target detection field: precision (Precision), recall (Recall), average precision mean (mAP@50, mAP@50:95), and model size (Model Size). These indicators not only measure the discrimination ability of the detection model for different defect categories, but also reflect its actual deployment value in industrial scenarios.

[0082] 1. Precision (Precision) Precision reflects the reliability of the model's prediction results, i.e., how many of the predicted positive samples are actually positive samples. The formula is as follows:

[0083] 2. Recall (Recall) Recall measures the detection ability of the model, i.e., how many of all positive samples are correctly identified. The formula is as follows:

[0084] Where FNFNFN represents the number of false negatives. A higher recall rate indicates that the model has a lower rate of missed detections when detecting small-scale or complex defects.

[0085] 3. Mean Average Precision (mAP) mAP is the most important evaluation indicator in object detection, which considers the balance between precision and recall. First, define the average precision (AP) as the area under the precision-recall curve:

[0086] Where P(R) represents the precision at different recall rates.

[0087] mAP@50: The average of each class AP when the IoU threshold is fixed at 0.5:

[0088] mAP@50-95 is the average of the average precision between the IoU threshold of 0.5 and 0.95, which is usually used to measure the comprehensive performance of the model under different IoU standards. Its calculation formula is:

[0089] 4. IoU (Intersection over Union) Calculation formula:

[0090] In the defect detection task, IoU is used to measure the overlap between the detected defect area and the true defect area.

[0091] This embodiment designs 5 groups of comparison models to verify the effect of each module. All models are tested under the same training parameters: 1. Base model: standard YOLOv8 without any improvement, as a performance benchmark, used to measure the improvement of subsequent modules.

[0092] 2. Model A: only add DRADPELAN module, focus on verifying the improvement of multi-scale feature extraction on defect detection accuracy; 3. Model B: only add DCFN module, focus on analyzing the optimization effect of multi-scale feature fusion on detection robustness in complex scenes; 4. Model C: only use lightweight design (C2f-DCFN+GSConvs+ VoVGSCSPC), verify the improvement of efficiency optimization on model size and inference speed, and observe the change of precision; 5.OURS: Fusion of DRADPELAN, MDPNet and lightweight design, embodying the synergistic effect of each module, verifying the performance of the complete scheme of "multi-scale extraction - feature fusion - lightweight".

[0093] The detection results are shown in Table 3 as follows: Table 3

[0094] As can be seen from Table 3, the three improved modules all positively contribute to the model performance, indicating that the synergistic effect of the three types of modules not only significantly improves the detection accuracy and robustness, but also takes into account the lightweight optimization, realizing the double improvement of performance and efficiency. The accuracy of the present invention is 92.4%, the recall rate is 90.6%, the mAP@50 reaches 0.938, and the mAP@50-95 reaches 0.568, which is the best overall performance. While ensuring lightweight, the performance is comprehensively improved, indicating that the improvement strategy proposed by the present invention is not only effective in theory, but also has good application prospect and popularization value in the actual detection task of train wheel tread defects.

[0095] In addition, by Figure 9 The mAP@50 convergence curves of each model in the training process are shown, which more directly reflects the influence of different improved modules on the training stability and final accuracy. The mAP@50 convergence curves of each model in the training process more directly reflect the influence of different improved modules on the training stability and final accuracy. The role of the present invention in reducing redundant calculation and improving reasoning efficiency is particularly suitable for resource-constrained industrial detection scenarios.

[0096] As Figure 10 shown, the model is visualized and analyzed by GradCAM++ heat map. Compared with the benchmark model, the model proposed by the present invention is more accurate in defect positioning, can clearly distinguish adjacent small defects, and effectively reduces the misjudgment of the background, directly verifying its excellent feature focusing ability and detection reliability.

[0097] Example 4 This example selects three types of mainstream models and compares them with the present invention: (1) YOLOv3-tiny (traditional lightweight) in Redmon J, Farhadi A. Yolov3: An incremental improvement[J]. arXiv preprint arXiv:1804.02767, 2018; (2) Deep-Sea Biological Detection Method Based on Lightweight YOLOv5n[J].Sensors, 2023, 23(20): DOI:10.3390 / S23208600, YOLOv5n (lightweight benchmark); (3) A Single-Stage Object Detection Framework for Industrial Applications. arXiv 2022, arXiv:2209.02976, YOLOv6s (high performance).

[0098] All models are tested under the same environment (Windows 10, Intel i5-8300 CPU, NVIDIA GTX 1050Ti GPU, CUDA 11.1) and parameters (initial learning rate 0.001, training 700 rounds) to ensure fairness.

[0099] The experimental results are shown in Table 4: Table 4

[0100] As can be seen from the table, YOLOv3-tiny, as an early lightweight model, has insufficient detection stability, especially difficult to adapt to small defects. As can be seen from the curve, it rises slowly and the final convergence value is low, and the overall performance is lagging behind. Although YOLOv5n finds a balance between lightweight and performance, the miss detection rate is high, and the curve fluctuates significantly in the later stage, which is difficult to meet the needs of high robustness scenarios. Although YOLOv6s is designed for performance, it does not meet expectations, and the model size is large and has the risk of overfitting, with a sharp curve fluctuation and poor deployment adaptability.

[0101] In comparison, the present application performs best in all core indicators while maintaining a lightweight size. As shown in Figure 11 , the corresponding curve rises quickly and converges smoothly without significant fluctuations, which not only reflects strong detection robustness for complex tread defects, but also has high deployment potential. As shown in Figure 12 , the model proposed by the present application performs excellently in wheel tread defect detection, can accurately identify typical defects such as dense point erosion and edge blur peeling, and has no repeated labeling problem, with confidence and positioning accuracy better than comparative models, effectively balancing detection performance and lightweight deployment needs.

[0102] In summary, the traditional and mainstream models have their own shortcomings, either lack of precision robustness, large missed defects, or high deployment difficulty. The present application considers both detection performance and inference efficiency, and fully verifies the effectiveness of the proposed method in wheel tread defect detection, laying a foundation for the promotion of complex scenes.

[0103] Example 5 In order to verify the robustness and generalization ability of the model proposed in the present application under different data distributions, the Northeast University (NEU) steel strip surface defect dataset

[31] is selected as the external test set to investigate the performance of the model on unseen industrial defect types, thereby further illustrating its generalization ability.

[0104] The NEU dataset was established by Song et al. and contains six typical surface defects of hot-rolled steel strips: inclusions (Inclusion, In), patches (Patch, Pa), cracks (Crack, Cr), pitted surfaces (Pitted surface, Ps), rolled-in scales (Rolled-in scale, RS), and scratches (Scratch, Sc). The dataset contains a total of 1800 grayscale images, 300 images for each type of defect. In this experiment, the dataset is divided into a training set and a test set in an 8:2 ratio, with 1440 images in the training set and 360 images in the test set, and both are trained under the same experimental environment. The experimental results are shown in Table 5 as follows: Table 5

[0105] Combining Table 5 and Figure 13 It can be seen that the performance of each model decreases compared to the wheel set dataset, but OURS has the smallest decline — mAP@50 (76.8%) and mAP@50-95 (42.4%) are the best, and the model is only 5.4MB; YOLOv3-tiny has the most serious decline (mAP@50 68.4%), YOLOv5 and YOLOv6 have highlights in precision / recall, but the curve fluctuates greatly.

[0106] The curve shows that OURS rises quickly (73% at 200 rounds) and stably (standard deviation 0.4%), YOLOv5 and YOLOv6 fluctuate significantly (standard deviation 0.9%-1.1%). This is because DRADPELAN extracts general defect features such as "edges, texture mutations", reducing dependence on single data; and OURS continues to have the advantage of lightweight, adapting to industrial equipment.

[0107] In summary, the present application has the best performance and the smallest decline in cross-dataset scenarios, verifying the universality and promotion value of the method.

[0108] Obviously, the above embodiments of the present application are merely exemplary but not intended to limit the embodiments of the present application. Based on the above description, any other variations or changes can be made by those skilled in the art without departing from the spirit and principles of the present application. It is not necessary to list all the embodiments here. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present application shall fall within the scope of the claims of the present application.

Claims

1. A defect lightweight detection model based on multi-scale dynamic fusion, characterized in that, The method comprises the following steps: A data collection module comprising a camera or a video camera collects defect images of a workpiece in different service states and different working conditions; A preprocessing module comprising an image denoising submodule, a geometric transformation submodule, a gray scale enhancement submodule, and an advanced data enhancement submodule denoises, geometrically transforms, and data enhances the collected data to construct diversified training samples; A multi-scale feature extraction backbone network is composed of four groups of multi-scale feature enhancement modules with different hole configurations, and the multi-scale feature enhancement module comprises a multi-scale hole reparameterization module and a hole pyramid attention module for adaptive feature extraction of defect scales; wherein The multi-scale hole reparameterization module comprises an original feature extraction branch and a deep feature extraction branch, adopts a double-path parallel structure, path one is an identity mapping branch which retains original feature information through 1*1 convolution; path two is a multi-scale feature extraction branch which is composed of a large kernel convolution and a plurality of hole convolutions with different hole rates in parallel to capture local details and context information in different ranges; The hole pyramid attention module is composed of a multi-scale hole convolution, a defect-sensitive activation function, and a channel attention fusion, the input features are first formed into multi-scale features through the multi-scale hole convolution, then the features of each branch are respectively subjected to nonlinear transformation through the defect-sensitive activation function to highlight the defect area and suppress the background, finally the channel attention mechanism is used to assign different channel weights to the feature maps of the three branches and perform weighted summation, and then feature integration and output are performed; An adaptive feature fusion module comprising a C2f-DCFN module optimized based on a dynamic mixed convolution module, a GSConvs module, and a VoVGSCSPC module, which deeply interacts and fuses the multi-scale features extracted above, retains multi-scale defect details, and reduces redundancy; A defect detection output module comprising three multi-scale detection head submodules, a classification branch, a regression branch, and a loss function calculation layer, which receives the features output by the adaptive feature fusion module, performs classification and bounding box regression, and finally outputs the category and position of the defect.

2. The defect lightweight detection model based on multi-scale dynamic fusion according to claim 1, wherein, The C2f-DCFN module is obtained by replacing the Bottleneck module of the original c2f module with a DCFN module, which uses DCFN to more comprehensively capture multi-scale context information, enhances context-dependent modeling and asymmetric feature coupling capability; The GSConvs module replaces the shuffle operation with a 1*1conv+LeakyReLU activation function based on the GSConv, learns the defect feature channel weight through training, dynamically enhances the key feature response strength, and improves the feature discriminability of small-scale microcracks and other defects; The VoVGSCSPC is obtained by stacking multiple GSBottleneck into a single GSBottleneckC module based on the VoVGSCSP, the hidden channel number is fixed at 50% of the output channel, and the GSBottleneckC uses DWConv instead of ordinary convolution as the shortcut branch.

3. An industrial defect lightweight detection method based on multi-scale dynamic fusion, characterized by the steps of The method comprises the following steps: S1. Collect image data, and pre-process the image data to obtain an input picture; S2. Perform multi-scale feature extraction on the input picture; S21. Perform dimension reduction processing on the input picture to obtain original features X1 and X2 from the channel dimension; S22. The original features X2 are respectively extracted by convolution, convolution kernel-dilated convolution to obtain shallow features and deep features F dil ; S23. The shallow features F and the deep features F dil are fused by weighting to obtain multi-scale features F , and the specific formula is: For output features, i.e. multi-scale features ; F std For shallow features, F dil For deep features denotes a weighted fusion S24. Forming different receptive fields by multi-scale hollow convolution Different receptive fields are formed by multi-scale hollow convolution, a defect-sensitive activation function is introduced again, the response difference between abnormal areas and background is regulated by channel threshold, and defect features are highlighted, wherein the defect-sensitive activation function is: In the formula, T e R^(Cxlxl) is a learnable channel threshold; a controls the negative response slope, and b controls the overall scaling. ReLU( ) represents a rectified linear unit. Finally, the channel attention is used to weight and fuse the branch features to output feature X3, which is expressed by the formula: wherein Y is an output feature, i.e., a fused output feature X3; a weight representing a branch; a learnable fusion weight, representing the output feature of each branch after defect activation; S25. Repeat steps S22-S24 to obtain fusion output feature X4 by taking the fusion output feature X3 as input; S26. The original features X1, X2, the fusion output features X3 and X4 are spliced to output an enhanced multi-scale feature P; S3. The multi-scale features P are first normalized and adapted, then up-sampled and aligned across scales for fusion, then enhanced by adaptive mixed convolution, and finally integrated to obtain discriminative features ; S31. Perform channel layer normalization on the multi-scale feature, perform channel adaptive transformation according to the cross-scale splicing requirement, and then perform spatial alignment and splicing fusion to obtain a cross-scale fusion feature; S32. split the cross-scale fusion feature into a main branch and an enhanced branch and the enhanced branch , the main branch is extracted by convolution and activation to obtain a feature ; S33. enhance branch First, the dynamic weight is generated, then it is extracted by multi-modal DW convolution branch, and finally the output features are fused , which is expressed as: ； wherein: Y is the output feature, is the corresponding weight for each group, is the DW convolution output; S34. The feature and the feature channel splicing and dimension unification to obtain the feature , and the feature and the feature residual connection to obtain the feature ; S35. The feature The GSConvs and VoVGSCSPC modules are integrated to obtain the discriminative features of the adaptive detection head ; S4. discriminative features output by S3 The detection result of wheel tread defects is obtained after the classified branch and the regression branch are combined and processed.

4. The industrial defect lightweight detection method based on multi-scale dynamic fusion according to claim 3, characterized in that, The pre-processing includes denoising, geometric transformation, gray scale enhancement and data enhancement.

5. The method according to claim 3, wherein, In the inference stage, the group of dilated convolutions is equivalent to a convolution by the reparameterization method, and the formula is as follows: ； wherein represents the output of the reasoning, i.e. the deep features F dil , Align( ) represents mapping different atrous convolution weights to a unified kernel size, represents the weight of the i-th branch, represents the mapped weight.

6. The industrial defect lightweight detection method based on multi-scale dynamic fusion of claim 3, wherein, The deep feature extraction weights and fuses the outputs of different branch dilated convolutions according to learnable weights during training, and the formula is expressed as: ， wherein: represents the output of the training phase, represents the weight of the i-th branch; * represents the convolution operation; learnable fusion weights, represents the output feature of the i-th branch.

7. The industrial defect lightweight detection method based on multi-scale dynamic fusion of claim 3, wherein, The normalization formula is: where X is the input feature, is the mean of the feature channel dimension, is the channel variance, .

8. The industrial defect lightweight detection method based on multi-scale dynamic fusion of claim 3, wherein, The dynamic weight is expressed as: ； is an input feature, i.e. the augmented branch , is a global average pooling, is a convolution operation, is a set of three weights.

9. The industrial defect lightweight detection method based on multi-scale dynamic fusion of claim 3, wherein, The specific steps of step S4 include: S41. Determine the discriminative features The input detection head, after feature enhancement and category prediction, obtains a defect category prediction probability. S42. Input the regression branch again, and obtain a defect candidate box through feature enhancement, parameter prediction and coordinate decoding; S43. Merge the obtained scale candidate boxes, and output the final defect detection result through confidence screening and non-maximum suppression.

10. The industrial defect lightweight detection method based on multi-scale dynamic fusion of claim 3, wherein, The method is applied to defect detection of a wheel tread.

Images