A bearing image processing and defect recognition method
By using a cascaded YOLO network and an attention mechanism, the problems of missed detection and false detection in the detection of minute defects on the surface of bearing rollers by traditional convolutional neural networks are solved, and high-precision defect identification is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUANGSHI BANGKE TECH CO LTD
- Filing Date
- 2026-01-15
- Publication Date
- 2026-06-05
AI Technical Summary
In complex environments, traditional convolutional neural networks struggle to effectively detect minute defects on the surface of bearing rollers, especially in complex contexts where they suffer from missed or false detections.
By adopting a cascaded YOLO network architecture and attention mechanism, and combining channel and spatial attention mechanisms, the detection capability of minute defects on the bearing roller surface is enhanced.
It improves the detection precision and accuracy of minute defects on the bearing roller surface, reduces the false detection rate under complex backgrounds, and meets the needs of large-scale industrial testing.
Smart Images
Figure CN122156049A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of bearing inspection, and specifically relates to a bearing image processing and defect identification method. Background Technology
[0002] As a core component of mechanical transmission systems, the performance and operational safety of bearings directly depend on the quality of their internal rollers. Even minute defects on the surface of these rollers can lead to accelerated wear, decreased precision, or even failure, ultimately causing equipment malfunctions. Therefore, defect detection of bearing rollers is a crucial step in ensuring the reliable operation of mechanical equipment. In traditional inspection scenarios, defect detection primarily relies on manual visual inspection, where inspectors visually examine the surface of the bearing rollers to determine the presence of defects. However, this manual inspection method has significant limitations. It is not only inefficient and ill-suited to the demands of large-scale industrial production, but also susceptible to human error. Visual fatigue, differences in experience, and subjective judgment bias among inspectors can all compromise the accuracy and consistency of inspection results, making it difficult to meet the quality control requirements of high-precision bearing manufacturing.
[0003] With the rapid development of computer vision technology, automatic defect detection methods based on image processing and machine learning are gradually replacing traditional manual inspection, becoming a research hotspot and development trend in the field of bearing roller quality inspection. These automatic inspection methods acquire high-definition images of the bearing roller surface using image acquisition equipment, preprocess and extract features from the images using image processing algorithms, and then combine these with machine learning models to classify and locate defects. This effectively overcomes the bottleneck of manual inspection, significantly improves inspection efficiency and accuracy, and reduces the impact of human factors on the inspection results, providing a feasible solution for large-scale industrial inspection.
[0004] Although deep learning-based object detection technology has made some progress in the detection of bearing roller defects, existing algorithms still have many problems that need to be solved. Taking the mainstream YOLO series object detection algorithms as an example, while they have excellent real-time detection performance and accuracy, and can meet the detection needs of most common scenarios, they still have performance shortcomings in complex detection environments. When the detection environment is complex, and there are complex textures on the surface of the bearing rollers, traditional convolutional neural network structures struggle to fully mine and capture the key features of tiny bearing roller defects, leading to increased false positives and missed detections, and a significant decrease in detection performance. Summary of the Invention
[0005] The present invention provides a bearing image processing and defect identification method, which can effectively solve the problems in the background art.
[0006] The present invention provides a bearing image processing and defect identification method, comprising the following steps:
[0007] S1. Build a bearing roller surface image acquisition platform to acquire images of bearing roller surface defect features;
[0008] S2. Perform grayscale conversion, normalization, and noise removal preprocessing on the acquired images;
[0009] S3. Cascade multiple YOLO detection modules in series to construct a cascaded YOLO network architecture, with the output of the previous YOLO detection module serving as the input of the next YOLO detection module;
[0010] S4. Select channel attention mechanism and spatial attention mechanism, embed them into the feature extraction stage of the cascaded YOLO network and between each YOLO detection module, and achieve organic integration with the cascaded YOLO network by adjusting parameters and connection methods;
[0011] S5. Construct a dataset of bearing roller surface images containing different defect types and complex conditions, expand the training data using data augmentation techniques, and train a cascaded YOLO network with an attention mechanism.
[0012] S6. Verify the performance of the trained algorithm through cross-validation and test set evaluation;
[0013] S7. Optimize the algorithm structure and parameters based on the verification results.
[0014] As a further optimization of the present invention, in the cascaded YOLO network architecture of step S3, the front-end YOLO detection module focuses on the preliminary screening of large-scale background areas, while the rear-end YOLO detection module focuses on the localization and identification of minute defects.
[0015] As a further optimization of the present invention, the step between S3 and S4 also includes the optimization of the cascaded YOLO network. The optimization includes adjusting the convolutional layer structure, selecting the activation function, and optimizing the loss function. The activation function adopts the Leaky ReLU function, and the loss function adopts a weighted loss function that combines position loss, confidence loss, and class loss.
[0016] As a further optimization of the present invention, in step S4, the selected channel attention mechanism and spatial attention mechanism adopt a squeeze-excitation structure to learn the importance weights of different channel features through squeeze-excitation operation; the spatial attention mechanism adopts the spatial branch structure of the convolutional attention module, first performing global pooling operation on the input features, and then generating a spatial attention map through convolution operation.
[0017] As a further optimization of the present invention, in step S4, the attention mechanism module is divided into two levels of embedding: the first level is embedded in the backbone feature extraction stage of the cascaded YOLO network; the second level is embedded between adjacent YOLO detection modules.
[0018] As a further optimization of the present invention, in step S4, the adjustment of parameters and connection methods includes: constructing a combined attention module using a serial connection mode, with channel attention first and spatial attention second, features first being filtered by channel weights and then spatial region focusing, and then adjusting the convolution kernel size, stride and fully connected layer parameters of the attention module through backpropagation training.
[0019] As a further optimization of the present invention, the data enhancement techniques in step S5 include at least three of the following: random cropping, rotation, flipping, brightness adjustment, and contrast adjustment, wherein the rotation angle range is -15° to 15°, and the cropping ratio of random cropping is 0.6 to 0.9 times that of the original image.
[0020] As a further optimization of the present invention, in step S5, during the network training process, the defect response intensity of the output features of the attention module is monitored in real time, and the embedding position weight of the attention module is dynamically fine-tuned according to the defect detection accuracy of the validation set.
[0021] As a further optimization of the present invention, the algorithm optimization in step S7 includes network pruning and quantization to meet real-time requirements and hardware resource constraints. The network pruning adopts a structured pruning strategy to cut off redundant convolutional kernels in the convolutional layer, and the quantization adopts 8-bit integer quantization.
[0022] As a further optimization of the present invention, step S7 includes adjusting the algorithm parameters and data preprocessing methods, extending to gear surface defect detection and blade surface defect detection, and achieving adaptation by adjusting the texture suppression parameters in image preprocessing and the anchor box size of the network.
[0023] This invention provides a bearing image processing and defect recognition method that combines a cascaded YOLO network with an attention mechanism for detecting defects on the surface of bearing rollers. The cascaded YOLO network, through the cascading of multi-stage detection modules, effectively addresses the insufficient detection accuracy of traditional YOLO algorithms in small target detection and complex backgrounds. Furthermore, the introduction of the attention mechanism further enhances the network's ability to focus on key features, enabling the algorithm to more accurately locate and identify minute defects and defect features under complex textures on the bearing roller surface. Attached Figure Description
[0024] Figure 1 This is a schematic diagram of the steps in this embodiment. Detailed Implementation
[0025] like Figure 1 As shown, this embodiment includes the following steps:
[0026] S1: Build a bearing roller surface image acquisition platform, that is, set up cameras and lights near the conveyor belt that transports bearing rollers. The lights illuminate the bearing rollers, and the cameras take pictures of the bearing rollers. The acquired pictures include images of surface defects of the bearing rollers.
[0027] The acquired images have preset resolution and sharpness, which meet the requirements for identifying minute defect features.
[0028] S2: The acquired images are preprocessed by grayscale conversion, normalization, and noise removal to eliminate interference factors in the images, enhance the distinguishability of defect features, and provide high-quality input data for subsequent feature extraction and detection algorithms.
[0029] Specifically, an adaptive filtering algorithm is used for noise removal.
[0030] Based on the YOLO series of target detection algorithms, multiple YOLO detection modules are cascaded to construct a cascaded YOLO network architecture. Each YOLO detection module focuses on target detection tasks of different scales or complexities, thereby improving the ability to detect small defects on the surface of bearing rollers and defects in complex backgrounds.
[0031] Specifically, multiple YOLO detection modules are cascaded in series, with the output of the preceding YOLO detection module serving as the input of the following YOLO detection module.
[0032] Traditional single YOLO networks have limited detection scale range, making it difficult to simultaneously handle large background areas and minute defects. In the cascaded architecture used in this embodiment, the front-end YOLO detection module first performs preliminary screening of the entire image or complex background, filtering out a large number of irrelevant background areas and narrowing the subsequent detection range; the rear-end YOLO detection module focuses on detecting minute defects in the narrowed area, and through progressively refined scale adaptation, it avoids the features of minute defects being overwhelmed by the features of complex backgrounds or large-scale targets.
[0033] In this embodiment, the front-end module learns basic background and defect contour features to provide more concise feature inputs for the back-end module. The back-end module focuses on learning fine features of minute defects, such as edge textures of minute cracks and grayscale differences of minute pits, which solves the problem of missed detection of minute defects caused by the excessive generalization and insufficient fineness of feature learning in traditional networks.
[0034] In complex backgrounds, traditional networks are prone to misidentifying background textures as defects. The cascaded architecture used in this embodiment reduces background interference in subsequent modules by performing preliminary filtering of complex backgrounds through front-end modules. At the same time, the continuous verification mechanism of multiple modules can reduce the probability of misjudgment by a single module. Only areas that have been identified as defects by multiple modules are finally output as results, effectively suppressing false detections caused by complex backgrounds.
[0035] In this embodiment, after step S3, there is a step of optimizing the cascaded YOLO network, which specifically includes adjusting the convolutional layer structure, selecting the activation function, and optimizing the loss function. The activation function used is the Leaky ReLU function, and the loss function is a weighted loss function that combines positional loss, confidence loss, and class loss.
[0036] S4. Select channel attention mechanism and spatial attention mechanism, and embed them into the feature extraction stage of the cascaded YOLO network or between each YOLO detection module. By adjusting the parameters and connection method, achieve organic integration with the cascaded YOLO network.
[0037] Specifically, it includes the following steps:
[0038] The first step is to select and adapt the appropriate model. Based on the characteristics of the bearing roller defects and the cascaded network, a combination scheme of channel attention mechanism and spatial attention mechanism is determined.
[0039] The channel attention mechanism focuses on feature channel weight allocation, employing a squeeze-excitation structure to learn the importance weights of different channel features through squeeze-excitation operations. The spatial attention mechanism focuses on defect region localization, using a spatial branching structure of a convolutional attention module. It first performs global pooling on the input features, then generates a spatial attention map through convolution operations, thereby focusing on defect-related regions. The channel and spatial attention mechanisms are complementary.
[0040] The second step involves precise design of the embedding location, dividing the attention mechanism module into two levels of embedding: the first level embeds the backbone feature extraction stage of the cascaded YOLO network, specifically after the bottleneck layer of the backbone network, to perform preliminary screening of global features and enhance the channel response and spatial focusing of defective basic features; the second level embeds the modules between adjacent YOLO detection modules, specifically after the feature mapping of the candidate image blocks output by the previous module and before the preprocessing of the input features of the subsequent module, to achieve secondary enhancement of the candidate region features and further suppress residual background interference.
[0041] The third step is to optimize the parameters and connection methods. A combined attention module is constructed using a serial connection mode, with channel attention preceding spatial attention, so that features are first filtered by channel weights and then focused on spatial regions. The convolutional kernel size, stride, and fully connected layer parameters of the attention module are adjusted through backpropagation training. At the same time, the feature dimension matching parameters of the attention module and the cascaded network are optimized to ensure that the feature dimension after attention processing is consistent with the input dimension of the subsequent network.
[0042] S5. Construct a dataset of bearing roller surface images containing different defect types and environmental conditions, expand the training data using data augmentation techniques, and train a cascaded YOLO network with an attention mechanism.
[0043] The data enhancement techniques include at least three of the following: random cropping, rotation, flipping, brightness adjustment, and contrast adjustment. The rotation angle ranges from -15° to 15°, and the cropping ratio of random cropping is 0.6 to 0.9 times that of the original image.
[0044] In this embodiment, during network training, the defect response intensity of the output features of the attention module is monitored in real time. Based on the defect detection accuracy of the validation set, the embedding position weight of the attention module is dynamically adjusted, i.e., the feature fusion ratio of the two-level embedding modules. Ultimately, the collaborative work of the attention mechanism and the cascaded YOLO network is realized, which can effectively enhance the recognition of small defect features and the anti-interference ability in complex backgrounds.
[0045] S6. Validate the performance of the trained algorithm through cross-validation and test set evaluation.
[0046] S7. Optimize the algorithm structure and parameters based on the verification results.
[0047] In this embodiment, the algorithm optimization involves network pruning and quantization to address real-time requirements and hardware resource limitations. The network pruning employs a structured pruning strategy to trim redundant convolutional kernels in the convolutional layers; quantization uses 8-bit integer quantization.
[0048] By adaptively adjusting the algorithm parameters and data preprocessing in this embodiment, it can be extended to other fields, such as gear surface defect detection and blade surface defect detection. Specifically, this can be achieved by adjusting the texture suppression parameters in image preprocessing and the anchor box size of the network.
[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.
Claims
1. A bearing image processing and defect identification method, characterized in that, Includes the following steps: S1. Build a bearing roller surface image acquisition platform to acquire images of bearing roller surface defect features; S2. Perform grayscale conversion, normalization, and noise removal preprocessing on the acquired images; S3. Cascade multiple YOLO detection modules in series to construct a cascaded YOLO network architecture, with the output of the previous YOLO detection module serving as the input of the next YOLO detection module; S4. Select channel attention mechanism and spatial attention mechanism, embed them into the feature extraction stage of the cascaded YOLO network and between each YOLO detection module, and achieve organic integration with the cascaded YOLO network by adjusting parameters and connection methods; S5. Construct a dataset of bearing roller surface images containing different defect types and complex conditions, expand the training data using data augmentation techniques, and train a cascaded YOLO network with an attention mechanism. S6. Verify the performance of the trained algorithm through cross-validation and test set evaluation; S7. Optimize the algorithm structure and parameters based on the verification results.
2. The bearing image processing and defect identification method according to claim 1, characterized in that, In the cascaded YOLO network architecture of step S3, the front-end YOLO detection module focuses on the initial screening of large-scale background areas, while the back-end YOLO detection module focuses on the localization and identification of minute defects.
3. The bearing image processing and defect identification method according to claim 1, characterized in that, Between steps S3 and S4, there is also optimization of the cascaded YOLO network. The optimization includes adjusting the convolutional layer structure, selecting the activation function, and optimizing the loss function. The activation function is the Leaky ReLU function, and the loss function is a weighted loss function that combines position loss, confidence loss, and class loss.
4. The bearing image processing and defect identification method according to claim 1, characterized in that, In step S4, among the selected channel attention mechanism and spatial attention mechanism, the channel attention mechanism adopts a squeeze-excitation structure, which learns the importance weights of different channel features through the squeeze-excitation operation; the spatial attention mechanism adopts the spatial branch structure of the convolutional attention module, which first performs global pooling operation on the input features, and then generates a spatial attention map through convolution operation.
5. The bearing image processing and defect identification method according to claim 1, characterized in that, In step S4, the attention mechanism module is divided into two levels of embedding: the first level embedding is the backbone feature extraction stage of the cascaded YOLO network; the second level embedding is between adjacent YOLO detection modules.
6. The bearing image processing and defect identification method according to claim 1, characterized in that, In step S4, the adjustment of parameters and connection methods includes: constructing a combined attention module using a serial connection mode, with channel attention first and spatial attention second. Features are first filtered by channel weights and then focused on spatial regions. Afterward, the kernel size, stride and fully connected layer parameters of the attention module are adjusted through backpropagation training.
7. The bearing image processing and defect identification method according to claim 1, characterized in that, The data augmentation techniques in step S5 include at least three of the following: random cropping, rotation, flipping, brightness adjustment, and contrast adjustment. The rotation angle ranges from -15° to 15°, and the cropping ratio of random cropping is 0.6 to 0.9 times that of the original image.
8. The bearing image processing and defect identification method according to claim 1, characterized in that, In step S5, during network training, the defect response intensity of the output features of the attention module is monitored in real time, and the embedding position weights of the attention module are dynamically fine-tuned based on the defect detection accuracy of the validation set.
9. The bearing image processing and defect identification method according to claim 1, characterized in that, The algorithm optimization in step S7 includes network pruning and quantization to meet real-time requirements and hardware resource constraints. The network pruning adopts a structured pruning strategy to cut redundant convolutional kernels in the convolutional layers, and the quantization adopts 8-bit integer quantization.
10. The bearing image processing and defect identification method according to claim 1, characterized in that, Step S7 also includes adjusting algorithm parameters and data preprocessing methods, extending to gear surface defect detection and blade surface defect detection, and achieving adaptation by adjusting texture suppression parameters in image preprocessing and the anchor box size of the network.