CN117541568A - Deep learning-based automobile brake disc surface defect detection method - Google Patents

Abstract

The invention relates to a method for detecting surface defects of an automobile brake disc based on deep learning. Comprising the following steps: 1. an image acquisition device is built to acquire an image of the surface of an automobile brake disc; 2. performing data preprocessing and marking on the obtained surface image of the automobile brake disc, and establishing an automobile brake disc surface defect data set; 3. constructing a model for detecting the surface defects of the automobile brake disc according to the surface defect data set of the automobile brake disc; 4. training the model by using the training set and the verification set and selecting an optimal detection model according to the evaluation index; 5. verifying the model by using the test set, and judging whether the performance of the model meets the requirement; 6. and detecting the surface of the automobile brake disc by using a defect detection model meeting the requirements, outputting a detection result, and realizing intelligent defect detection and identification. The invention can improve the detection efficiency and accuracy of the surface defects of the automobile brake disc, reduce the labor cost, and can be rapidly adapted to the detection of the surface defects of novel products, thereby shortening the development period.

Description

A method for detecting surface defects of automobile brake discs based on deep learning

Technical Field

The present invention belongs to the technical field of automobile brake discs, and specifically is a method for detecting surface defects of automobile brake discs based on deep learning.

Background Art

As one of the important components in the automobile braking system, automobile brake discs convert the kinetic energy of the car into heat energy through friction with other brake components, thereby achieving deceleration or stopping of the car. In the production process of automobile brake discs, due to the influence of relevant production factors such as production materials, equipment technology and manual operation, it is inevitable that sand holes, scratches and dirt defects will appear on the surface of automobile brake discs. If a car is equipped with a defective brake disc, the defect will cause damage to the brake disc during use. The car will have a longer deceleration distance at the least, and the car's brakes will fail at the worst, which is very likely to cause traffic accidents. Therefore, it is necessary to conduct strict surface defect inspection on automobile brake discs before leaving the factory to ensure their braking effect and prevent problematic brake discs from entering the market at the source.

At present, the surface defect detection of automobile brake discs before leaving the factory usually adopts manual inspection and machine vision inspection. Among them, manual inspection often has problems such as high cost, poor real-time performance, false detection, and serious missed detection, which cannot meet the high-precision requirements of automobile brake disc surface defect detection; and surface defect detection based on machine vision, due to the complex factory production environment and the influence of objective factors such as noise during image acquisition, often requires multiple image processing algorithms to cooperate for feature extraction, resulting in high complexity of defect detection algorithms, which is not conducive to the subsequent maintenance and improvement of staff. In addition, the traditional image processing method has limited feature extraction capabilities, and the feature extraction effect is not sufficient when facing complex defect information, which leads to poor detection effect and weak robustness. In recent years, with the continuous development of deep learning technology in the field of vision, a large number of researchers have applied deep learning-based visual inspection technology to the field of industrial inspection, thereby avoiding the shortcomings of manual inspection and alleviating the problem of poor robustness of industrial inspection based on machine vision.

Summary of the invention

The present invention provides a method for detecting surface defects of automobile brake discs based on deep learning. On the basis of the original YOLOv5l target detection model, the backbone network, feature fusion network and prediction network are improved to adapt to the automobile brake disc surface defect dataset, thereby improving the detection accuracy and efficiency, enhancing the robustness of the detection algorithm, realizing the automatic detection of automobile brake discs, and solving the problems of low efficiency and poor accuracy of manual detection.

The technical solution of the present invention is described as follows in conjunction with the accompanying drawings:

A method for detecting surface defects of automobile brake discs based on deep learning, comprising the following steps:

Step 1: Build an image acquisition device at the end of the production line to obtain the surface images of automobile brake discs under different lighting conditions;

Step 2: pre-process and annotate the acquired automobile brake disc surface images, establish an automobile brake disc surface defect dataset, and divide all annotated sample images into a training set, a validation set, and a test set;

Step 3: Based on the automobile brake disc surface defect dataset, a surface defect detection model for automobile brake disc is constructed;

Step 4: Use the samples of the training set and the samples of the validation set to iteratively train and validate the defect detection model established in step 3, and select the weighted model with the highest mAP index in the validation set as the optimal automobile brake disc surface defect detection model;

Step 5: Use the samples of the test set in step 2 to test the optimal automobile brake disc surface defect detection model obtained in step 4, and evaluate whether the detection model meets the accuracy requirements according to the mAP index; if it does not meet the requirements, repeat step 4 and continue to train the detection model; if it meets the requirements, execute step 6;

Step 6: Use a defect detection model that meets the accuracy requirements to inspect the surface of the automobile brake disc and output the test results, thereby realizing intelligent defect detection and identification.

Furthermore, the specific method of step one is as follows:

The image acquisition device composed of a high-resolution CMOS area array industrial camera and a surface light source is used to photograph the automobile brake disc under different lighting conditions;

The different lighting conditions include lighting intensity and lighting quality.

Furthermore, the specific method of step 2 is as follows:

21) The original image of the automobile brake disc surface is cropped to obtain sub-images of size 640×640. There is a 20% overlap between the two sub-images. Each sub-image is mirrored and flipped with a probability of 50%;

22) Use LabelImg annotation software to annotate the image processed in step 21) and output it in VOC annotation format;

23) Convert the VOC annotation format to the COCO annotation format to complete the construction of the automobile brake disc surface defect dataset;

24) The automobile brake disc surface defect dataset constructed in step 23) is divided into a training set, a validation set and a test set.

Furthermore, the division ratio of the training set, validation set and test set is 8:1:1.

Furthermore, the specific method of step three is as follows:

31) Using the YOLOv5l detection model as the benchmark model, the self-built CSPDarkFormer backbone feature extraction network is used to replace the CSPDarkNet53 backbone feature extraction network in the original YOLOv5l;

32) The original feature fusion network of YOLOv5l is improved as the feature fusion network of the automobile brake disc surface defect detection model.

33) Decouple the prediction network in the original YOLOv5l, remove the prediction confidence branch, retain the category prediction branch and the position prediction branch, and build a loss function based on the anchor-free box detection mechanism for model training.

Further, the specific method of step 31) is as follows:

The CSPDarkFormer backbone feature extraction network is based on the Transformer architecture, and uses a convolutional module to replace the common self-attention module in the Transformer architecture. It includes a 5-layer network structure, which is: the first layer is the Stem layer, the 2nd to 4th layers are composed of the Patch Embedding layer and the DFM module, and the 5th layer is composed of the Patch Embedding layer, the DFM module and the SPPF module. The feature maps of the 3rd to 5th layers of the backbone feature extraction network are output as the input of the improved feature fusion network, which are denoted as {C ₃ ,C ₄ ,C ₅ } respectively;

Among them, the Patch Embedding layer includes a convolution module with a size of 3 and a stride of 2, and a BN normalization layer;

The DFM module includes a C2Former module and a multi-layer perceptron module MLP;

The feature map is first normalized, then processed by the C2Former module and added to itself element by element to form a residual connection, and then processed by the normalization operation and the multi-layer perceptron MLP module to achieve information interaction between different channels;

The calculation process of the DFM module is as follows:

Where X _DFM is the output feature map of the DFM module; X is the feature map after downsampling by the Patch Embedding layer, that is, Input is the input feature map of the Patch Embedding layer. is a 3×3 convolution operation with a step size of 2 for the input feature map; Norm ₁ (·) and Norm ₂ (·) are BN normalization; C2Former (·) is the C2Former module; MLP (·) is the MLP multi-layer perceptron module; F ₁ is the intermediate feature map of the DFM module;

The C2Former module includes two CBS modules with a convolution kernel size of 1×1 and several DarkBlock modules; wherein the CBS module includes a convolution module, a BN normalization layer and a SiLU activation function;

The calculation process of the C2Former module is as follows:

X _C2Former =f ^1×1 ([[n×D(f ^1×1 (X))],X])

Where X _C2Former is the output feature map of the C2Former module; X is the input feature map of the C2Former module; n is the number of DarkBlock modules, which are connected in series, and each feature map is concatenated after being processed by a DarkBlock module, and finally concatenated with the input X; f ^1×1 is a CBS module with a convolution kernel size of 1×1; [·] is the Concate concatenation operator; D is a DarkBlock module, and the calculation process is as follows:

Where X _DarkBlock is the output feature map of the DarkBlock module; X is the input feature map of the DarkBlock module; is a 5×5 DCBS module, which includes a 5×5 depthwise separable convolution module, a BN normalization layer, and a SiLU activation function; f ^3×3 is a CBS module with a convolution kernel size of 3×3; the DarkBlock module uses a residual connection method;

The multi-layer perceptron MLP module includes a convolution module with a convolution kernel size of 1×1 and a SiLU activation function. The calculation process is as follows:

X _MLP =g ^1×1 (τ(g ^1×1 (X))

Where _XMLP is the output feature map of the MLP module; X is the input feature map of the MLP module; τ(·) is the SiLU activation function; g1 ^×1 is the convolution module with a convolution kernel size of 1×1.

Further, the specific method of step 32) is as follows:

First, the SimSE lightweight channel attention module is added to the input part, and then the C3 module in the original YOLOv5l feature fusion network is replaced by the C2Former module. Finally, the three-layer feature maps of 80×80, 40×40, and 20×20 are output as the input of the prediction network, denoted as {P ₃ ,P ₄ ,P ₅ }, where the DarkBlock module in the C2Former module has no residual connection;

The SimSE lightweight channel attention module is a lightweight SE attention module, and the calculation process is as follows:

X _SimSE = σ(g ^1×1 (Avg(X))

Where _XSimSE is the output feature map of the SimSE lightweight channel attention module; X is the input feature map of the SimSE lightweight channel attention module; Avg(·) is the global average pooling operation; g ^1×1 is the convolution module with a convolution kernel size of 1×1; σ(·) is the HardSigmoid activation function;

Further, the specific method of step 33) is as follows:

The loss function of the prediction network is specifically expressed as:

Loss＝λ _cls L _cls +λ _bbox L _bbox +λ _dfl L _dfl

Where λ _cls , λ _bbox and λ _dfl are the weight coefficients of each loss, which are selected as 0.5, 7.5 and 0.375 respectively; L _cls is the classification loss; L _bbox and L _dfl together constitute the positioning loss;

The main positioning loss function L _bbox is CIOU loss, which is specifically expressed as:

Where IOU is the intersection-over-union ratio between the predicted box and the real box; ρ ² (b, b ^gt ) is the L2 distance between the midpoint of the predicted box and the midpoint of the real box; c is the longest diagonal distance between the predicted box and the real box; α and v are used to measure the aspect ratio, which is specifically expressed as:

Where w and h are the width and height of the predicted box, respectively, and w ^gt and h ^gt are the width and height of the real box, respectively;

The auxiliary positioning loss function L _dfl is Distribution Focal Loss, which is specifically expressed as:

L _dfl =-((y _i+1 -y)log(S _i )+(yy _i )log(S _i+1 ))

Where y is the distance from the center point of the real box to the boundary; _yi and yi ₊₁ are the integer values of the real value y rounded down and rounded up respectively; _Si and Si ₊₁ are the probability values from the center point of the predicted box to the boundary; log is the logarithmic operation;

The L _cls is the cross entropy loss, expressed as:

Where n is the number of samples; is the category probability score of the predicted sample; _yi is the true label.

Furthermore, in step 4,

The training method includes: the total number of training rounds is 600 rounds, the batch size is 16, the optimizer is the AdamW optimizer, the initial learning rate is 0.001, the cos learning rate decay method is adopted, and the random cropping, Mosaic, Mixup, and random HSV transformation data enhancement methods are used during the training process. Mosaic and MixUp data enhancements are turned off in the last 100 rounds;

The training process includes: the backbone feature extraction network inputs a training image with a resolution of 640×640, first passes through the Stem layer to reduce the feature map resolution to twice the original, then passes through 3 feature extraction layers composed of a Padding Embedding layer and a DFM module in sequence, and finally passes through a feature extraction layer composed of a Padding Embedding layer, a DFM module and an SPPF module to complete the final feature extraction. Finally, the backbone feature extraction network outputs three layers of feature maps with resolutions of 80×80, 40×40, and 20×20, respectively, denoted as {C ₃ ,C ₄ ,C ₅ }, as the input feature maps of the feature fusion network; after the feature fusion network obtains {C ₃ ,C ₄ ,C ₅ }, it first passes through the SimSE lightweight channel attention module, and then passes through the improved feature fusion network, and finally obtains three layers of feature maps with resolutions of 80×80, 40×40, and 20×20, respectively, denoted as {P ₃ ,P ₄ ,P ₅ }, as the input feature map of the prediction network;

The prediction network consists of three layers, and the output resolution of each layer corresponds to the input resolution. Each layer includes two branches, which predict the category and position respectively. Each branch contains two CBS modules with a convolution kernel size of 3×3 and a convolution with a convolution kernel size of 1×1. Among them, the prediction network branch with an output resolution of 80×80 is responsible for predicting small targets; the prediction network branch with an output resolution of 40×40 is responsible for predicting medium targets; the prediction network branch with an output resolution of 20×20 is responsible for predicting large targets. Finally, the automobile brake disc defect detection model outputs the defect category and location information, which are marked by a rectangular box.

Furthermore, in step 5, mAP is used as an evaluation indicator, which is specifically defined as follows:

Where n is the number of categories; AP is the area enclosed by the P-R curve and the coordinate axis, which represents the prediction accuracy of each type of defect; mAP is the average accuracy of each type of defect, where P is the precision and R is the recall, which is specifically defined as:

In the formula, TP is the number of correctly predicted positive samples; FP is the number of correctly predicted negative samples; FN is the number of incorrectly predicted positive samples.

The beneficial effects of the present invention are:

1) The CSPDarkFormer backbone feature extraction network constructed by the present invention is based on the Transformer architecture. The convolution module is used to replace the self-attention module commonly used in the Transformer architecture, which retains the cross-channel information interaction capability in the Transformer structure and significantly increases the model's ability to capture local context information. At the same time, it alleviates the deficiency of the self-attention mechanism in the Transformer on large feature maps, which leads to a significant decrease in detection speed due to excessive calculation, thereby improving the efficiency and accuracy of feature extraction;

2) The improved feature fusion network of the present invention adopts a large convolution kernel for feature fusion and feature extraction, which expands the receptive field of the model and improves the detection accuracy of the model;

3) Based on the original prediction network, the present invention decouples the prediction network and adopts an anchor-free detection mechanism to prevent the influence of prior information such as anchor boxes on the detection accuracy, thereby improving the bounding box regression accuracy;

4) Based on the anchor-free detection mechanism, the present invention optimizes the loss function, jointly represents the DFL loss and the bounding box regression loss, increases the comprehensiveness of positive and negative sample training, improves the convergence speed of bounding box regression, and further improves the model detection accuracy and efficiency.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

Fig. 1 is a schematic diagram of a process of the present invention;

Fig. 2a is an example diagram of sand holes;

Figure 2b is a schematic diagram of a scratch;

Figure 2c is a schematic diagram of dirt;

FIG3 is a schematic diagram of the structure of a surface defect detection model for automobile brake discs based on deep learning;

FIG4 is a schematic diagram of a DFM module;

Figure 5 is a schematic diagram of the C2Former module;

Figure 6a is a CBS module;

Figure 6b is a schematic diagram of a DCBS module;

Figure 7 is a diagram showing the experimental detection results.

DETAILED DESCRIPTION

The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only parts related to the present invention, rather than all structures, are shown in the accompanying drawings.

Embodiment 1

Referring to FIG. 1 , this embodiment provides a method for detecting surface defects of automobile brake discs based on deep learning, comprising the following steps:

Step 1: Build an image acquisition device at the end of the production line to obtain the surface image of the automobile brake disc under different lighting conditions, as follows:

The image acquisition device composed of a high-resolution CMOS area array industrial camera and a surface light source is used to photograph the automobile brake disc under different lighting conditions;

The different lighting conditions include adjusting the lighting intensity and lighting quality to simulate a complex environment that meets factory production conditions.

Step 2: Preprocess and annotate the acquired automobile brake disc surface images, establish an automobile brake disc surface defect dataset, and divide all annotated sample images into training set, validation set and test set, as follows:

21) The original image of the automobile brake disc surface is cropped to obtain sub-images of size 640×640. There is a 20% overlap between the two sub-images. Each sub-image is mirrored and flipped with a probability of 50%;

22) Use LabelImg annotation software to annotate the image processed in step 21) and output it in VOC annotation format;

23) Convert the VOC annotation format to COCO annotation format to facilitate model testing and complete the construction of the automobile brake disc surface defect dataset;

24) The automobile brake disc surface defect dataset constructed in step 23) is divided into a training set, a validation set and a test set; the division ratio of the training set, the validation set and the test set is 8:1:1.

Step 3: Based on the automobile brake disc surface defect dataset, a car brake disc surface defect detection model is constructed. The YOLOv5l detection model is used as the benchmark model. According to the characteristics of the automobile brake disc surface defects, the backbone feature extraction network, feature fusion network and prediction network of the YOLOv5l detection model are optimized as follows:

31) Using the YOLOv5l detection model as the benchmark model, the self-built CSPDarkFormer backbone feature extraction network was used to replace the CSPDarkNet53 backbone feature extraction network in the original YOLOv5l;

The CSPDarkFormer backbone feature extraction network is based on the Transformer architecture, and uses a convolutional module to replace the common self-attention module in the Transformer architecture. It includes a 5-layer network structure, which is: the first layer is the Stem layer, the 2nd to 4th layers are composed of the Patch Embedding layer and the DFM module, and the 5th layer is composed of the Patch Embedding layer, the DFM module and the SPPF module. The feature maps of the 3rd to 5th layers of the backbone feature extraction network are output as the input of the improved feature fusion network, which are denoted as {C ₃ ,C ₄ ,C ₅ } respectively;

Among them, the Patch Embedding layer includes a convolution module with a size of 3 and a stride of 2, and a BN normalization layer;

The DFM module mainly includes a C2Former module and a multi-layer perceptron module MLP;

The feature map is first normalized, then processed by the C2Former module and added to itself element by element to form a residual connection, and then processed by the normalization operation and the multi-layer perceptron MLP module to achieve information interaction between different channels;

The calculation process of the DFM module is as follows:

Where X _DFM is the output feature map of the DFM module; X is the feature map after downsampling by the Patch Embedding layer, that is, Input is the input feature map of the Patch Embedding layer. is a 3×3 convolution operation with a step size of 2 for the input feature map; Norm ₁ (·) and Norm ₂ (·) are BN normalization; C2Former (·) is the C2Former module; MLP (·) is the MLP multi-layer perceptron module; F ₁ is the intermediate feature map of the DFM module;

The C2Former module includes two CBS modules with a convolution kernel size of 1×1 and several DarkBlock modules; wherein the CBS module includes a convolution module, a BN normalization layer and a SiLU activation function;

The calculation process of the C2Former module is as follows:

X _C2Former =f ^1×1 ([[n×D(f ^1×1 (X))],X])

Where X _C2Former is the output feature map of the C2Former module; X is the input feature map of the C2Former module; n is the number of DarkBlock modules, which are connected in series, and each feature map is concatenated after being processed by a DarkBlock module, and finally concatenated with the input X; f ^1×1 is a CBS module with a convolution kernel size of 1×1; [·] is the Concate concatenation operator; D is a DarkBlock module, and the calculation process is as follows:

Where X _DarkBlock is the output feature map of the DarkBlock module; X is the input feature map of the DarkBlock module; is a 5×5 DCBS module, which includes a 5×5 depthwise separable convolution module, a BN normalization layer and a SiLU activation function; f ^3×3 is a CBS module with a convolution kernel size of 3×3; in the backbone feature extraction network, in order to alleviate the gradient vanishing phenomenon, the residual connection method is used in the DarkBlock module;

The multi-layer perceptron MLP module includes a convolution module with a convolution kernel size of 1×1 and a SiLU activation function. The calculation process is as follows:

X _MLP =g ^1×1 (τ(g ^1×1 (X))

Where _XMLP is the output feature map of the MLP module; X is the input feature map of the MLP module; τ(·) is the SiLU activation function; g1 ^×1 is the convolution module with a convolution kernel size of 1×1.

32) Improve the original feature fusion network of YOLOv5l as the feature fusion network of the automobile brake disc surface defect detection model;

First, add the SimSE lightweight channel attention module to the input part, then replace the C3 module in the feature fusion network in the original YOLOv5l with the C2Former module, and finally output the three-layer feature maps of 80×80, 40×40, and 20×20 as the input of the prediction network, denoted as {P ₃ ,P ₄ ,P ₅ }; taking the P ₄ output layer as an example, the specific calculation process can be expressed as:

In the formula, is the output feature map of _P4 layer; and are the intermediate feature maps of the 3rd, 4th and 5th layers respectively; C ₄ is the output feature map of the 4th layer of the backbone feature extraction network; UP(·) and Down(·) are upsampling and downsampling operations respectively; f ^1×1 is a CBS module with a convolution kernel size of 1×1; [·] represents the Concate concatenation operator; C2Former(·) is a C2Fomer module, which has the same residual connection method as the C3 module in the YOLOv5l feature fusion network, that is, the DarkBlock module in the C2Former module has no residual connection.

The SimSE lightweight channel attention module is a lightweight SE attention module, and the calculation process is as follows:

X _SimSE = σ(g ^1×1 (Avg(X))

Where _XSimSE is the output feature map of the SimSE lightweight channel attention module; X is the input feature map of the SimSE lightweight channel attention module; Avg(·) is the global average pooling operation; g ^1×1 is the convolution module with a convolution kernel size of 1×1; σ(·) is the HardSigmoid activation function;

33) Decouple the prediction network in the original YOLOv5l, remove the prediction confidence branch, retain the category prediction branch and the position prediction branch, and build a loss function based on the anchor-free box detection mechanism for model training.

The loss function of the prediction network is specifically expressed as:

Loss＝λ _cls L _cls +λ _bbox L _bbox +λ _dfl L _dfl

Where λ _cls , λ _bbox and λ _dfl are the weight coefficients of each loss, which are selected as 0.5, 7.5 and 0.375 respectively; L _cls is the classification loss; L _bbox and L _dfl together constitute the positioning loss;

The main positioning loss function L _bbox is CIOU loss, which is specifically expressed as:

Where IOU is the intersection-over-union ratio between the predicted box and the real box; ρ ² (b, b ^gt ) is the L2 distance between the midpoint of the predicted box and the midpoint of the real box; c is the longest diagonal distance between the predicted box and the real box; α and v are used to measure the aspect ratio, which is specifically expressed as:

Where w and h are the width and height of the predicted box, respectively, and w ^gt and h ^gt are the width and height of the real box, respectively;

The auxiliary positioning loss function L _dfl is the Distribution Focal Loss, which is jointly expressed with the positioning loss function to improve the convergence speed of bounding box regression. It is specifically expressed as:

L _dfl =-((y _i+1 -y)log(S _i )+(yy _i )log(S _i+1 ))

Where y is the distance from the center point of the real box to the boundary; _yi and yi ₊₁ are the integer values of the real value y rounded down and rounded up respectively; _Si and Si ₊₁ are the probability values from the center point of the predicted box to the boundary; log is the logarithmic operation;

The L _cls is the cross entropy loss, expressed as:

Where n is the number of samples; is the category probability score of the predicted sample; _yi is the true label.

Step 4: Use the samples of the training set and the samples of the validation set to iteratively train and validate the defect detection model established in step 3, and select the weighted model with the highest mAP index in the validation set as the optimal automobile brake disc surface defect detection model;

The training method includes: the total number of training rounds is 600 rounds, the batch size is 16, the optimizer is the AdamW optimizer, the initial learning rate is 0.001, the cos learning rate decay method is adopted, and the random cropping, Mosaic, Mixup, and random HSV transformation data enhancement methods are used during the training process. Mosaic and MixUp data enhancements are turned off in the last 100 rounds;

The training process includes: the backbone feature extraction network inputs a training image with a resolution of 640×640, first passes through the Stem layer to reduce the feature map resolution to twice the original, then passes through 3 feature extraction layers composed of a Padding Embedding layer and a DFM module in sequence, and finally passes through a feature extraction layer composed of a Padding Embedding layer, a DFM module and an SPPF module to complete the final feature extraction. Finally, the backbone feature extraction network outputs three layers of feature maps with resolutions of 80×80, 40×40, and 20×20, respectively, denoted as {C ₃ ,C ₄ ,C ₅ }, as the input feature maps of the feature fusion network; after the feature fusion network obtains {C ₃ ,C ₄ ,C ₅ }, it first passes through the SimSE lightweight channel attention module, and then passes through the improved feature fusion network, and finally obtains three layers of feature maps with resolutions of 80×80, 40×40, and 20×20, respectively, denoted as {P ₃ ,P ₄ ,P ₅ }, as the input feature map of the prediction network;

The prediction network consists of three layers, and the output resolution of each layer corresponds to the input resolution. Each layer includes two branches, which predict the category and position respectively. Each branch contains two CBS modules with a convolution kernel size of 3×3 and a convolution with a convolution kernel size of 1×1. Among them, the prediction network branch with an output resolution of 80×80 is responsible for predicting small targets; the prediction network branch with an output resolution of 40×40 is responsible for predicting medium targets; the prediction network branch with an output resolution of 20×20 is responsible for predicting large targets; finally, the automobile brake disc defect detection model outputs the defect category and location information, both marked by rectangular boxes.

Step 5: Use the samples of the test set in step 2 to test the optimal automobile brake disc surface defect detection model obtained in step 4, and evaluate whether the detection model meets the accuracy requirements according to the mAP index; if it does not meet the requirements, repeat step 4 and continue to train the detection model; if it meets the requirements, execute step 6;

In step 5, mAP is used as the evaluation indicator, which is specifically defined as follows:

Where n is the number of categories; AP is the area enclosed by the P-R curve and the coordinate axis, which represents the prediction accuracy of each type of defect; mAP is the average accuracy of each type of defect, where P is the precision and R is the recall, which is specifically defined as:

In the formula, TP is the number of correctly predicted positive samples; FP is the number of correctly predicted negative samples; FN is the number of incorrectly predicted positive samples.

Step 6: Use a defect detection model that meets the accuracy requirements to inspect the surface of the automobile brake disc and output the test results, thereby realizing intelligent defect detection and identification.

Embodiment 2

Referring to FIG. 1 , this embodiment takes a practical situation as an example to detect the surface defects of the automobile brake disc through the first embodiment, as follows:

S1. Build an image acquisition device at the end of the production line to obtain the surface images of automobile brake discs under different lighting conditions.

Specifically, a high-resolution CMOS industrial area array camera and a surface light source are used to form an image acquisition device to capture images of the surface of automobile brake discs on the assembly line. During the acquisition process, the brightness and illumination angle of the light source need to be adjusted to obtain surface defect images with different light intensities and qualities to simulate the complex production environment of the factory and enrich the data set. The defects mainly include three categories: sand holes, scratches and dirt. The resolution of each image is 5472×3648.

S2. Preprocess and annotate the automobile brake disc surface images, establish an automobile brake disc surface defect dataset, and divide all annotated sample images into training set, verification set and test set.

Specifically, since the original image resolution is too large, it cannot be combined with the subsequent defect detection algorithm, so the original image needs to be cropped first, where the size of the cropped sub-image is 640×640. During the cropping process, a 20% overlap area needs to be retained between each sub-image. After the cropping is completed, the cropped image is cleaned to remove image data that does not contain defects or has unclear defects. After the data cleaning is completed, the data set is uniformly mirrored and flipped with a probability of 50% to expand the data set. The LabelImg image annotation tool is used to annotate the image of the automobile brake disc surface defect data set, output the VOC annotation format, and then convert the VOC annotation format to the COCO annotation format to facilitate the quantification of the model performance. The construction of the automobile brake disc surface defect data set is completed, and finally the data set is divided into a training set, a validation set, and a test set according to the ratio of 8:1:1. Some sample example images are shown in Figures 2a, 2b, and 2c, where Figure 2a is a sand hole defect, Figure 2b is a scratch defect, and Figure 2c is a dirt defect.

S3. Based on the automobile brake disc surface defect dataset, a automobile brake disc surface defect detection model is constructed.

Specifically, taking the YOLOv5l detection model as the benchmark model, according to the automobile brake disc surface defect dataset, the YOLOv5l target detection model is optimized in three aspects: backbone feature extraction network, feature fusion network, and prediction network. The overall framework diagram is shown in Figure 3. The specific optimization scheme is as follows:

(1) Backbone feature extraction network;

The backbone feature extraction network is based on the Transformer architecture. The convolution module is used to replace the common self-attention module in the Transformer architecture. The CSPDarkFormer backbone feature extraction network is built and replaces the original CSPDarkNet53 backbone feature extraction network. The CSPDarkFormer backbone feature extraction network includes a 5-layer network structure, which is: the first layer is the Stem layer, the 2nd to 4th layers are composed of the Patch Embedding layer and the DFM module, and the 5th layer is composed of the PatchEmbedding layer, the DFM module and the SPPF module. The feature maps of the 3rd to 5th layers of the backbone feature extraction network are output as the input of the improved feature fusion network, which are denoted as {C ₃ ,C ₄ ,C ₅ } respectively;

The Patch Embedding layer includes a convolution module of size 3 and stride 2 and a BN normalization layer, which is mainly used for feature map downsampling. When given input X, its calculation process is as follows:

PE(X)＝Norm(Conv(X))

Where PE(·) is the Patch Embedding layer; X is the input feature map of the Patch Embedding layer; Conv(·) is the convolution operation; Norm(·) is the BN normalization, which means that the input x is subtracted from the mean of the current batch, then divided by the standard deviation of the current batch, and finally multiplied by γ and added to β to obtain the normalized _xi value, which is expressed as:

In the formula, x is the input feature map; ε is a minimum value that is not 0 to prevent the divisor from being 0, γ and β are the learnable parameters of the model; E(x) and Var(x) are the mean and variance of the current batch, respectively, which can be expressed as:

Where n is the current batch size.

The DFM module is an important feature extraction component in the backbone feature extraction network, as shown in FIG4 , and mainly includes: a C2Former module and a multi-layer perceptron module MLP. The feature map is first normalized, then processed by the C2Former module and added element by element to itself to form a residual connection, and then normalized and multi-layer perceptrons are used to realize information interaction between different channels. The calculation process of the DFM module is as follows:

Where X _DFM is the output feature map of the DFM module; X is the feature map after downsampling by the Patch Embedding layer; Norm ₁ (·) and Norm ₂ (·) are BN normalizations; F ₁ is the intermediate feature map; C2Former (·) represents the C2Former module; MLP (·) is the MLP multi-layer perceptron module;

The C2Former module, as shown in FIG5 , includes two CBS modules with a convolution kernel size of 1×1 and several DarkBlock modules. When the input is X, the CBS module with a convolution kernel size of 1×1 is first used to preliminarily extract features and double the number of channels. Then, the channels are split into two, and the first part continues to extract features. During the feature extraction process, a Concate operation is required after each DarkBlock module is processed. Then, the second part is concatenated with the feature map after feature extraction of the first part. Finally, the number of output channels is adjusted by the CBS module with a convolution kernel size of 1×1.

Among them, the feature extraction part adopts the DarkBlock module, which includes a CBS module with a convolution kernel size of 3×3 and a DCBS module with a convolution kernel size of 5×5, which are connected in series to form a DarkBlock module. At the same time, a residual connection is added to the module to prevent the gradient from disappearing.

The multi-layer perceptron module MLP is the basic module in the Transformer architecture, which is mainly used for information interaction between channels. Its calculation process is as follows:

X _MLP =g ^1×1 (τ(g ^1×1 (X))

Where X _MLP is the output feature map of the MLP module; X is the input feature map of the MLP module; τ(·) is the SiLU activation function; g is the convolution module with a ^1×1 convolution kernel size of 1×1;

The CSPDarkFormer backbone feature extraction network constructed in the present invention is based on the Transformer architecture and uses a convolutional module to replace the commonly used self-attention module in the Transformer structure, thereby retaining the cross-channel information interaction capability of the Transformer and significantly increasing the ability of the model step to obtain local contextual information. At the same time, it alleviates the deficiency of the self-attention mechanism in the Transformer that the detection speed is greatly reduced due to excessive computation on large feature maps, thereby improving the efficiency and accuracy of feature extraction.

(2) Feature fusion network;

The improved feature fusion network is used as the feature fusion network in the automobile brake disc surface defect detection model. Compared with the original YOLOv5l feature fusion network, the improvement is as follows: first, the SimSE lightweight channel attention module is added to the input part, and then the C3 module of the feature fusion network in YOLOv5l is replaced by the C2Former module. Finally, the three-layer feature maps of 80×80, 40×40, and 20×20 are output as the input of the prediction network, respectively, and are recorded as {P ₃ ,P ₄ ,P ₅ }. Taking the P ₄ output layer as an example, the specific calculation process can be expressed as:

In the formula, is the output feature map of _P4 layer; and are the intermediate feature maps of the 3rd, 4th and 5th layers respectively; C ₄ is the output feature map of the 4th layer of the backbone feature extraction network; UP(·) and Down(·) represent upsampling and downsampling operations respectively, f ^1×1 represents the CBS module with a convolution kernel size of 1×1, [·] represents the Concate concatenation operator, and C2Former(·) is the C2Former module, which has the same residual connection method as the C3 module in the YOLOv5l feature fusion network, that is, the DarkBlock module of the C2Former module in the feature fusion network has no residual connection.

The lightweight channel attention module SimSE is a lightweight improvement of the SE attention module. The original SE attention mechanism uses a fully connected layer to redistribute channel weights. When the input feature map is large, the fully connected layer has a large amount of parameters and calculations, which is not conducive to the actual deployment of the model. Therefore, in the SimSE channel attention mechanism module, the number of fully connected layers is appropriately reduced and convolution is used to replace the fully connected layer to reduce the amount of calculation and parameters. The calculation process is as follows:

X _SimSE = σ(g ^1×1 (Avg(X))

Where _XSimSE is the output feature map of the SimSE lightweight channel attention module; X is the input of the channel attention module; Avg(·) represents the global average pooling layer; g is the convolution module with a ^1×1 convolution kernel size of 1×1; σ(·) is the HardSigmoid activation function.

The improved feature fusion network of the present invention adopts a large convolution kernel for feature fusion and feature extraction, which expands the receptive field of the model and improves the detection accuracy of the model.

(3) Prediction network;

The prediction network is decoupled based on the original YOLOv5l prediction network, removing the prediction confidence branch, retaining the category prediction branch and the position prediction branch, and building a loss function based on the anchor-free box detection mechanism for model training. The loss function of the automobile brake disc surface defect detection model is as follows:

Loss＝λ _cls L _cls +λ _bbox L _bbox +λ _dfl L _dfl

Where λ _cls , λ _bbox and λ _dfl are the loss weight coefficients, which are selected as 0.5, 7.5 and 0.375 respectively.

L _cls is the classification loss, using the cross entropy loss function, which can be specifically expressed as:

Where n is the number of samples; is the category probability score of the predicted sample; _yi is the true label; log is the logarithmic operation.

L _bbox and L _dfl together constitute the positioning loss, where L _bbox is the CIOU loss function and L _dfl is the DistributionFocal Loss loss function, which can be specifically expressed as:

Where IOU is the intersection-over-union ratio between the predicted box and the true box; ρ ² (b, b ^gt ) is the L2 distance between the midpoint of the predicted box and the midpoint of the true box; c is the longest diagonal distance between the predicted box and the true box; α and v are used to measure the aspect ratio, which can be specifically expressed as:

Where w and h are the width and height of the predicted box, respectively, and w ^gt and h ^gt are the width and height of the real box, respectively.

L _dfl =-((y _i+1 -y)log(S _i )+(yy _i )log(S _i+1 ))

Where y is the distance from the center point of the true box to the boundary; _yi and yi ₊₁ represent the integer values of the true value y rounded down and rounded up respectively; _Si and Si ₊₁ represent the probability values from the center point of the predicted box to the boundary.

Based on the original prediction network, the present invention decouples the prediction network and adopts an anchor-free detection mechanism to prevent the influence of prior information such as anchor boxes on the detection accuracy, thereby improving the bounding box regression accuracy. In addition, based on the anchor-free detection mechanism, the loss function is optimized, and the DFL loss is jointly represented with the bounding box regression CIOU loss, which increases the comprehensiveness of positive and negative sample training, improves the convergence speed of bounding box regression, and further improves the model detection accuracy and efficiency.

S4. Use the training set samples and the validation set samples to iteratively train and validate the defect detection model established in S3, and select the weighted model with the highest mAP index in the validation set as the final model.

Specifically, the training method includes: the total number of training rounds is 600 rounds, the batch size is 16, the optimizer is the AdamW optimizer, the initial learning rate is 0.001, the cos learning rate decay method is adopted, and random cropping, Mosaic, Mixup, and random HSV transformation data enhancement methods are used during training. In the last 100 rounds, Mosaic and MixUp data enhancements are turned off to prevent overfitting.

The training process includes: the backbone feature extraction network inputs a training image with a resolution of 640×640, first passes through the Stem layer to reduce the feature map resolution to twice the original, then passes through 3 feature extraction layers composed of a Padding Embedding layer and a DFM module in sequence, and finally passes through a feature extraction layer composed of a Padding Embedding layer, a DFM module and an SPPF module to complete the final feature extraction. Finally, the backbone feature extraction network outputs three layers of feature maps with resolutions of 80×80, 40×40, and 20×20, respectively, denoted as {C ₃ ,C ₄ ,C ₅ }, as the input feature maps of the feature fusion network; after the feature fusion network obtains {C ₃ ,C ₄ ,C ₅ }, it first passes through the SimSE lightweight channel attention module, and then passes through the improved feature fusion network, and finally obtains three layers of feature maps with resolutions of 80×80, 40×40, and 20×20, respectively, denoted as {P ₃ ,P ₄ ,P ₅ }, as the input feature map of the prediction network; the prediction network contains three layers, the output resolution of each layer corresponds to the input resolution, each layer includes two branches, predicting the category and position respectively, each branch contains two CBS modules with a convolution kernel size of 3×3 and a convolution kernel size of 1×1, among which the prediction network branch with an output resolution of 80×80 is responsible for predicting small targets, such as sand holes and small dirt or scratch defects, the prediction network branch with an output resolution of 40×40 is responsible for predicting medium targets, and the prediction network branch with an output resolution of 20×20 is responsible for predicting large targets, such as large-area dirt or long scratch defects. Finally, the automobile brake disc defect detection model outputs defect category and location information, both marked by rectangular boxes.

The CBS module and DCBS module used in the present invention are shown in Figures 6a and 6b respectively, wherein the CBS module includes a convolution module, a BN normalization layer and a SiLU activation function, and the DCBS includes a depth-separable convolution module composed of separable convolution and point convolution, a BN normalization layer and a SiLU activation function.

The activation functions used in the present invention are SiLU and Hardsigmoid, which can be specifically expressed as:

SiLU(x)＝x×sigmoid(x)

S5, using the test set samples in S2 to test the optimal automobile brake disc surface defect detection model obtained in S4, and evaluating whether the detection model meets the requirements according to the mAP index; if it does not meet the requirements, repeat S4 and continue to train the model; if it meets the requirements, execute S6;

Specifically, mAP is used as the evaluation indicator, and its specific definition is as follows:

Where n is the number of categories, AP is the area enclosed by the P-R curve and the coordinate axis, which represents the prediction accuracy of each type of defect; mAP is the average accuracy of each type of defect, where P is the precision and R is the recall, which can be specifically defined as:

In the formula, TP is the number of correctly predicted positive samples; FP is the number of correctly predicted negative samples; FN is the number of incorrectly predicted positive samples

S6. Use the defect detection model that meets the requirements to inspect the surface of the automobile brake disc, output the test results, and realize intelligent defect detection and identification.

Experimental verification

Using the automobile brake disc surface defect dataset established by S4 in Example 2, a comparative experiment was conducted on the automobile brake disc surface defect detection model, YOLOv5l, YOLOv7l, Faster R-CNN and RetinaNet target detection models. The results are shown in Table 1.

Table 1

As can be seen from Table 1, the surface defect detection model for automobile brake discs proposed in the present invention has a good detection effect, which is 1.5% and 3% higher than YOLOv5l and YOLOv7l respectively, 8.1% higher than RetinaNet, and 8.3% higher than the two-stage target detection model Faster R-CNN, meeting the actual production accuracy requirements. Some experimental results are shown in Figure 7, where the value in the upper left corner of each picture is the probability value of this type of defect.

In summary, based on the original YOLOv5l target detection model, the present invention improves the backbone network, feature fusion network and prediction network to adapt to the automobile brake disc surface defect dataset, thereby improving the detection accuracy and efficiency, enhancing the robustness of the detection algorithm, and realizing the automatic detection of automobile brake discs.

Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims (10)

1. A method for detecting surface defects of automobile brake discs based on deep learning, characterized in that it comprises the following steps: Step 1: Build an image acquisition device at the end of the production line to obtain the surface images of automobile brake discs under different lighting conditions; Step 2: pre-process and annotate the acquired automobile brake disc surface images, establish an automobile brake disc surface defect dataset, and divide all annotated sample images into a training set, a validation set, and a test set; Step 3: Based on the automobile brake disc surface defect dataset, a surface defect detection model for automobile brake disc is constructed; Step 4: Use the samples of the training set and the samples of the validation set to iteratively train and validate the defect detection model established in step 3, and select the weighted model with the highest mAP index in the validation set as the optimal automobile brake disc surface defect detection model; Step 5: Use the samples of the test set in step 2 to test the optimal automobile brake disc surface defect detection model obtained in step 4, and evaluate whether the detection model meets the accuracy requirements according to the mAP index; if it does not meet the requirements, repeat step 4 and continue to train the detection model; if it meets the requirements, execute step 6; Step 6: Use a defect detection model that meets the accuracy requirements to inspect the surface of the automobile brake disc and output the test results, thereby realizing intelligent defect detection and identification. 2. According to the method for detecting surface defects of automobile brake discs based on deep learning in claim 1, it is characterized in that the specific method of step 1 is as follows: The image acquisition device composed of a high-resolution CMOS area array industrial camera and a surface light source is used to photograph the automobile brake disc under different lighting conditions; The different lighting conditions include lighting intensity and lighting quality. 3. According to the method for detecting surface defects of automobile brake discs based on deep learning in claim 1, it is characterized in that the specific method of step 2 is as follows: 21) The original image of the automobile brake disc surface is cropped to obtain sub-images of size 640×640. There is a 20% overlap between the two sub-images. Each sub-image is mirrored and flipped with a probability of 50%; 22) Use LabelImg annotation software to annotate the image processed in step 21) and output it in VOC annotation format; 23) Convert the VOC annotation format to the COCO annotation format to complete the construction of the automobile brake disc surface defect dataset; 24) The automobile brake disc surface defect dataset constructed in step 23) is divided into a training set, a validation set and a test set. 4. According to the deep learning-based automobile brake disc surface defect detection method of claim 3, it is characterized in that the division ratio of the training set, the validation set and the test set is 8:1:1. 5. According to the method for detecting surface defects of automobile brake discs based on deep learning in claim 1, it is characterized in that the specific method of step three is as follows: 31) Using the YOLOv5l detection model as the benchmark model, the self-built CSPDarkFormer backbone feature extraction network was used to replace the CSPDarkNet53 backbone feature extraction network in the original YOLOv5l; 32) Improve the original feature fusion network of YOLOv5l as the feature fusion network of the automobile brake disc surface defect detection model; 33) Decouple the prediction network in the original YOLOv5l, remove the prediction confidence branch, retain the category prediction branch and the position prediction branch, and build a loss function based on the anchor-free box detection mechanism for model training. 6. According to the method for detecting surface defects of automobile brake discs based on deep learning in claim 5, it is characterized in that the specific method of step 31) is as follows: The CSPDarkFormer backbone feature extraction network includes a 5-layer network structure, which is: the first layer is the Stem layer, the 2nd to 4th layers are composed of the Patch Embedding layer and the DFM module, and the 5th layer is composed of the Patch Embedding layer, the DFM module and the SPPF module. The feature maps of the 3rd to 5th layers of the backbone feature extraction network are output as the input of the bidirectional feature pyramid network, which are respectively denoted as {C ₃ ,C ₄ ,C ₅ }; Among them, the Patch Embedding layer includes a convolution module with a size of 3 and a stride of 2, and a BN normalization layer; The DFM module mainly includes a C2Former module and a multi-layer perceptron module MLP; The feature map is first normalized, then processed by the C2Former module and added to itself element by element to form a residual connection, and then normalized and multi-layer perceptron MLP module is used to realize information interaction between different channels; The calculation process of the DFM module is as follows: Where X _DFM is the output feature map of the DFM module; X is the feature map after downsampling by the Patch Embedding layer, that is, Input is the input feature map of the Patch Embedding layer. is a 3×3 convolution operation with a step size of 2 for the input feature map; Norm ₁ (·) and Norm ₂ (·) are BN normalization; C2Former (·) is the C2Former module; MLP (·) is the MLP multi-layer perceptron module; F ₁ is the intermediate feature map of the DFM module; The C2Former module includes two CBS modules with a convolution kernel size of 1×1 and several DarkBlock modules; wherein the CBS module includes a convolution module, a BN normalization layer and a SiLU activation function; The calculation process of the C2Former module is as follows: X _C2Former =f ^1×1 ([[n×D(f ^1×1 (X))],X]) Where X _C2Former is the output feature map of the C2Former module; X is the input feature map of the C2Former module; n is the number of DarkBlock modules, which are connected in series, and each feature map is concatenated after being processed by a DarkBlock module, and finally concatenated with the input X; f ^1×1 is a CBS module with a convolution kernel size of 1×1; [·] is the Concate concatenation operator; D is a DarkBlock module, and the calculation process is as follows: Where X _DarkBlock is the output feature map of the DarkBlock module; X is the input feature map of the DarkBlock module; is a 5×5 DCBS module, which includes a 5×5 depthwise separable convolution module, a BN normalization layer, and a SiLU activation function; f ^3×3 is a CBS module with a convolution kernel size of 3×3; the DarkBlock module uses a residual connection method; The multi-layer perceptron MLP module includes a convolution module with a convolution kernel size of 1×1 and a SiLU activation function. The calculation process is as follows: X _MLP =g ^1×1 (τ(g ^1×1 (X)) Where _XMLP is the output feature map of the MLP module; X is the input feature map of the MLP module; τ(·) is the SiLU activation function; g1 ^×1 is the convolution module with a convolution kernel size of 1×1. 7. According to the method for detecting surface defects of automobile brake discs based on deep learning in claim 5, it is characterized in that the specific method of step 32) is as follows: First, the SimSE lightweight channel attention module is added to the input part, and then the C3 module in the original YOLOv5l feature fusion network is replaced by the C2Former module. Finally, the three-layer feature maps of 80×80, 40×40, and 20×20 are output as the input of the prediction network, denoted as {P ₃ ,P ₄ ,P ₅ }, where the DarkBlock module in the C2Former module has no residual connection; The SimSE lightweight channel attention module is a lightweight SE attention module, and the calculation process is as follows: X _SimSE = σ(g ^1×1 (Avg(X)) Where _XSimSE is the output feature map of the SimSE lightweight channel attention module; X is the input feature map of the SimSE lightweight channel attention module; Avg(·) is the global average pooling operation; g1 ^×1 is the convolution module with a convolution kernel size of 1×1; σ(·) is the HardSigmoid activation function. 8. According to the method for detecting surface defects of automobile brake discs based on deep learning in claim 5, it is characterized in that the specific method of step 33) is as follows: The loss function of the prediction network is specifically expressed as: Loss＝λ _cls L _cls +λ _bbox L _bbox +λ _dfl L _dfl Where λ _cls , λ _bbox and λ _dfl are the weight coefficients of each loss, which are selected as 0.5, 7.5 and 0.375 respectively; L _cls is the classification loss; L _bbox and L _dfl together constitute the positioning loss; The main positioning loss function L _bbox is CIOU loss, which is specifically expressed as: Where IOU is the intersection-over-union ratio between the predicted box and the real box; ρ ² (b, b ^gt ) is the L2 distance between the midpoint of the predicted box and the midpoint of the real box; c is the longest diagonal distance between the predicted box and the real box; α and v are used to measure the aspect ratio, which is specifically expressed as: Where w and h are the width and height of the predicted box, respectively, and w ^gt and h ^gt are the width and height of the real box, respectively; The auxiliary positioning loss function L _dfl is Distribution Focal Loss, which is specifically expressed as: L _dfl =-((y _i+1 -y)log(S _i )+(yy _i )log(S _i+1 )) Where y is the distance from the center point of the real box to the boundary; _yi and yi ₊₁ are the integer values of the real value y rounded down and rounded up respectively; _Si and Si ₊₁ are the probability values from the center point of the predicted box to the boundary; log is the logarithmic operation; The L _cls is the cross entropy loss, expressed as: Where n is the number of samples; is the category probability score of the predicted sample; _yi is the true label. 9. According to the method for detecting surface defects of automobile brake discs based on deep learning in claim 5, it is characterized in that in the step 4, The training method includes: the total number of training rounds is 600 rounds, the batch size is 16, the optimizer is the AdamW optimizer, the initial learning rate is 0.001, the cos learning rate decay method is adopted, and the random cropping, Mosaic, Mixup, and random HSV transformation data enhancement methods are used during the training process. Mosaic and MixUp data enhancements are turned off in the last 100 rounds; The training process includes: the backbone feature extraction network inputs a training image with a resolution of 640×640, first passes through the Stem layer to reduce the feature map resolution to twice the original, then passes through 3 feature extraction layers composed of a Padding Embedding layer and a DFM module in sequence, and finally passes through a feature extraction layer composed of a Padding Embedding layer, a DFM module and an SPPF module to complete the final feature extraction. Finally, the backbone feature extraction network outputs three layers of feature maps with resolutions of 80×80, 40×40, and 20×20, respectively, denoted as {C ₃ ,C ₄ ,C ₅ }, as the input feature maps of the feature fusion network; after the feature fusion network obtains {C ₃ ,C ₄ ,C ₅ }, it first passes through the SimSE lightweight channel attention module, and then passes through the improved feature fusion network, and finally obtains three layers of feature maps with resolutions of 80×80, 40×40, and 20×20, respectively, denoted as {P ₃ ,P ₄ ,P ₅ }, as the input feature map of the prediction network; The prediction network consists of three layers, and the output resolution of each layer corresponds to the input resolution. Each layer includes two branches, which predict the category and position respectively. Each branch contains two CBS modules with a convolution kernel size of 3×3 and a convolution with a convolution kernel size of 1×1. Among them, the prediction network branch with an output resolution of 80×80 is responsible for predicting small targets; the prediction network branch with an output resolution of 40×40 is responsible for predicting medium targets; the prediction network branch with an output resolution of 20×20 is responsible for predicting large targets; finally, the automobile brake disc defect detection model outputs the defect category and location information, both marked by rectangular boxes. 10. According to the method for detecting surface defects of automobile brake discs based on deep learning in claim 5, it is characterized in that in the step 5, mAP is used as an evaluation index, which is specifically defined as follows: Where n is the number of categories; AP is the area enclosed by the P-R curve and the coordinate axis, which represents the prediction accuracy of each type of defect; mAP is the average accuracy of each type of defect, where P is the precision and R is the recall, which is specifically defined as: In the formula, TP is the number of correctly predicted positive samples; FP is the number of correctly predicted negative samples; FN is the number of incorrectly predicted positive samples.