Light-weight aluminum material flaw detection method suitable for industrial environment

Abstract

The invention discloses a light-weight aluminum material flaw detection method suitable for an industrial environment. The method comprises the following steps: acquiring a related labeled flaw image from an aluminum material flaw data set; performing data enhancement processing on the image; constructing a light-weight aluminum material defect detection network, wherein the network comprises a light-weight backbone network, a light-weight multi-scale feature fusion network and a light-weight decoupling detection head; training the light-weight aluminum material flaw detection network by adopting an aluminum material flaw data set to obtain a light-weight aluminum material flaw detection model suitable for an industrial environment; the model is used for detecting the produced aluminum material; and identifying possible flaw parts. The method can be used for detecting various flaws on the surface of the aluminum material, has a better effect on tiny flaws and flaws difficult to identify by naked eyes, is low in calculation overhead requirement on a detection element, can better adapt to an industrial environment, and provides powerful support for aluminum material production.

Description

Technical Field

[0001] The present invention relates to the technical field of image detection and target recognition, and in particular to a lightweight aluminum material defect detection method suitable for an industrial environment. Background Art

[0002] With the continuous development and improvement of deep learning technology, its application in industrial production has become an important way to improve product quality and production efficiency. Especially in the field of aluminum manufacturing, since aluminum plays a key role in multiple industries, it is crucial to ensure the high-quality production of aluminum.

[0003] Although the existing aluminum defect detection technology has made some progress, it still faces some technical challenges in practical applications. For example, some detection schemes may require complex hardware configuration and a large amount of computing resources, which not only increases the overall cost of the system, but also limits its adaptability in different industrial environments. In addition, for certain types of defects, such as minor scratches or local color changes, existing technologies may find it difficult to achieve ideal detection accuracy. Summary of the invention

[0004] In response to the above-mentioned technical problems, this technical solution provides a lightweight aluminum defect detection method suitable for industrial environments. Building a lightweight aluminum defect detection network can reduce hardware configuration requirements and computational complexity while ensuring detection accuracy; it can not only effectively reduce the cost investment in the detection process, but also improve detection efficiency; it can effectively solve the above-mentioned problems.

[0005] The present invention is achieved through the following technical solutions: A lightweight aluminum material defect detection method suitable for industrial environments, comprising the steps of: Step 1: Obtain relevant labeled defect images from the aluminum defect dataset. The images contain defective parts and non-defective parts. Step 2: Perform data enhancement on the image; Step 3: Build a lightweight aluminum defect detection network, which includes a lightweight backbone network, a lightweight multi-scale feature fusion network, and a lightweight decoupling detection head; The lightweight backbone network is composed of a slice convolution module, four lightweight detection modules and a spatial pyramid pooling module in series; the lightweight detection module is composed of a lightweight convolution ghostconv and three lightweight feature processing modules c2f-star in series; Step 4: Use the aluminum defect dataset to train a lightweight aluminum defect detection network to obtain a lightweight aluminum defect detection model suitable for industrial environments; Step 5: Use the lightweight aluminum material defect detection model obtained in step 4 to detect the produced aluminum materials; Step 6: Identify possible defects in the produced aluminum.

[0006] Furthermore, the lightweight convolution ghostconv described in step three is implemented by the following method: first, ordinary convolution is used to reduce the number of channels to half of the original number, and the obtained feature map is called the intrinsic feature map. Then, deep convolution is used to extract features for each channel of the intrinsic feature map, and the extracted feature map is called the ghost feature map. Finally, the intrinsic feature map and the ghost feature map are added to obtain the output feature map.

[0007] Furthermore, the lightweight feature processing module c2f-star described in step three is implemented by the following method: first, the number of output channels is multiplied by the expansion rate to obtain the number of hidden layer channels, and then the number of channels of the input feature map is transformed into twice the number of hidden layer channels through 1×1 convolution, and then the channels of the feature map are evenly divided into two parts, one of which is directly passed to the output, and the other part is input to the starblock module for processing; specifically, this part of the data first passes through a deep convolution layer with a convolution kernel size of 7, and then is divided into two paths, which are respectively increased in dimension through two fully connected layers f1 and f2, and the output of f1 is activated using the ReLU6 activation function, and the output result of the ReLU6 activation is multiplied element by element with the output result of f2; the result of the multiplication is reduced to the dimension of the input through the fully connected layer g; then the output of the fully connected layer g is processed again using deep convolution, and the processing result is added to the input of the starblock module to obtain the final output.

[0008] Furthermore, the lightweight multi-scale feature fusion network described in step three is composed of four feature fusion modules connected in series; the first two feature fusion modules in the four feature fusion modules are composed of an upsampling module, a feature map connection module and a lightweight feature processing module c2f-star connected in series; the last two feature fusion modules are composed of a convolution module, a feature map connection module and a lightweight feature processing module c2f-star connected in series; the feature map connection modules in the four feature fusion modules are respectively connected to the corresponding module input and the output of the 6th, 4th, 12th and 9th layer models.

[0009] Furthermore, the lightweight decoupled detection head described in step three is implemented by the following method: first, two serially connected group convolutions are used for feature extraction to achieve parameter sharing; then, the feature maps obtained by the group convolution are output to a 1×1 convolution layer for bounding box prediction and a 1×1 convolution layer for category prediction, respectively, and finally the bounding box loss and category loss are obtained for back propagation.

[0010] Furthermore, the defects described in step one include dents, non-conductivity, scratches, orange peel, leakage, bruises, pits, convex powder, coating cracks and dirty spots.

[0011] Furthermore, in the data enhancement processing of the image described in step 2, the method of data expansion of the image mainly includes: adding image noise, adjusting image brightness, image shearing, image rotation, image translation and mirroring, and finally expanding the data set to five times the original size. Beneficial Effects

[0012] The invention proposes a lightweight aluminum material defect detection method suitable for industrial environments, which has the following beneficial effects compared with the prior art: The present invention adopts a backbone network composed of a lightweight convolution ghostconv and three lightweight feature processing modules c2f-star stacked in series. Ghostconv first generates an intrinsic feature map with a small number of channels through a small number of standard convolutions to aggregate inter-channel information, and then uses grouped convolutions performed on the intrinsic feature map one by one, and finally splices the two groups of feature maps to generate a new feature map. Since the computational overhead of grouped convolution is much smaller than that of ordinary convolution, the network can learn more useful feature representations without significantly increasing the computational complexity. C2f-star utilizes the StarOperation operation, which performs element-level multiplication operations in a low-dimensional space, so that the model can map the input to a high-dimensional, nonlinear feature space without increasing the computational complexity. This allows c2f-star to produce richer feature representations at a lower computational cost.

[0013] The present invention adopts a lightweight multi-scale feature fusion network composed of four feature fusion modules connected in series. The first two feature fusion modules of the four feature fusion modules are composed of an upsampling module, a feature map connection module and a lightweight feature processing module c2f-star connected in series, and the last two feature fusion modules are composed of a convolution module, a feature map connection module and a lightweight feature processing module c2f-star connected in series. The feature map connection modules in the four feature fusion modules are respectively connected to the corresponding module input and the output of the 6th, 4th, 12th and 9th layer models.

[0014] The present invention adopts a lightweight decoupled detection head, which realizes parameter sharing in the two tasks of category prediction and bounding box prediction, and changes ordinary convolution into group convolution, which greatly reduces the computational overhead of the detection head in the feature extraction process. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0016] Figure 2 It is a schematic diagram of the ghostConv structure in the present invention.

[0017] Figure 3 It is a schematic diagram of the C2f-Star structure in the present invention.

[0018] Figure 4 It is a schematic diagram of the structure of the lightweight decoupling detection head in the present invention.

[0019] Figure 5 It is a schematic diagram of the network structure of the present invention.

[0020] Figure 6 It is a schematic diagram of the detection results of the present invention. DETAILED DESCRIPTION

[0021] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Under the premise of not departing from the design concept of the present invention, various modifications and improvements made by ordinary persons in the art to the technical solutions of the present invention should all fall within the protection scope of the present invention. Example 1

[0022] like Figure 1 As shown, a lightweight aluminum material defect detection method suitable for industrial environments includes the following steps: Step 1: Get the image of the aluminum material to be tested.

[0023] Relevant labeled defect images are obtained from the aluminum defect dataset as samples. The defects in the images include: dents, non-conductivity, scratches, orange peel, leakage, bruises, pits, convex powder, coating cracks, and dirty spots. The samples include images with defective parts and images without defective parts.

[0024] Step 2: Perform data enhancement on the image.

[0025] Since the amount of image data to be tested is small, directly using the original data for model training will result in low model generalization performance and poor data fitting. Therefore, it is necessary to enhance and expand the images in the aluminum defect dataset to improve the generalization performance of the model. The enhancement methods mainly include: adding image noise, adjusting image brightness, image shearing, image rotation, image translation and mirroring, and finally expanding the dataset to five times the original size.

[0026] Step 3: Build a lightweight aluminum defect detection network. The detection network includes a lightweight backbone network, a lightweight multi-scale feature fusion network and a lightweight decoupling detection head.

[0027] The lightweight backbone network consists of a slice convolution module, four lightweight detection modules, and a spatial pyramid pooling module in series. The lightweight detection module consists of a lightweight convolution ghostconv and three lightweight feature processing modules c2f-star in series.

[0028] The lightweight convolution ghostconv is implemented by the following method: first, ordinary convolution is used to reduce the number of channels to half of the original number, and the obtained feature map is called the intrinsic feature map. Then, deep convolution is used to extract features for each channel of the intrinsic feature map. The extracted feature map is called the ghost feature map. Finally, the intrinsic feature map and the ghost feature map are added to obtain the output feature map.

[0029] The lightweight feature processing module c2f-star is implemented by the following method: first, the number of output channels is multiplied by the expansion rate to obtain the number of hidden layer channels, and then the number of channels of the input feature map is transformed into twice the number of hidden layer channels through 1×1 convolution. Then, the channels of the feature map are evenly divided into two parts, one of which is directly passed to the output, and the other is input to the starblock module for processing; specifically, this part of the data first passes through a deep convolution layer with a convolution kernel size of 7, and then is divided into two paths, which are respectively increased in dimension through two fully connected layers f1 and f2. The output of f1 is activated using the ReLU6 activation function, and the output result of the ReLU6 activation is multiplied element by element with the output result of f2; the result of the multiplication is reduced to the dimension of the input through the fully connected layer g; then the output of the fully connected layer g is processed again using deep convolution, and the processing result is added to the input of the starblock module to obtain the final output.

[0030] The lightweight multi-scale feature fusion network is composed of four feature fusion modules connected in series; the first two feature fusion modules of the four feature fusion modules are composed of an upsampling module, a feature map connection module and a lightweight feature processing module c2f-star connected in series; the last two feature fusion modules are composed of a convolution module, a feature map connection module and a lightweight feature processing module c2f-star connected in series; the feature map connection modules in the four feature fusion modules are respectively connected to the corresponding module input and the output of the 6th, 4th, 12th and 9th layer models.

[0031] The lightweight decoupled detection head is implemented by the following method: first, two serially connected group convolutions are used for feature extraction to achieve parameter sharing; then, the feature maps obtained by the group convolution are output to a 1×1 convolution layer for bounding box prediction and a 1×1 convolution layer for category prediction, respectively, and finally the bounding box loss and category loss are obtained for back propagation.

[0032] Step 4: Use the aluminum defect dataset to train the lightweight aluminum defect detection network to obtain a lightweight aluminum defect detection model suitable for industrial environments.

[0033] Step 5: Use the lightweight aluminum defect detection model obtained in step 4 to detect the produced aluminum.

[0034] Step 6: Identify possible defects in the produced aluminum.

[0035] In order to verify the feasibility and superiority of this embodiment, the inventor conducted a comparative experiment for verification.

[0036] In this experiment, Yolov8 is first run as the benchmark model for effect comparison, and then three innovative improvements, c2f_star, ghostconv and lightweight decoupling detection head, are added in turn to verify the improvement of the detection accuracy and lightweight effect of the model by the improvements.

[0037] The specific improvement data in this embodiment is shown in the following table, where MAP@0.5 represents the average precision mean, which can comprehensively reflect the detection accuracy of the model, GFLOPS represents the floating point operation amount, 1GFLOPS means one billion floating point operations per second, parameters represents the number of parameters, and FPS represents the number of frames per second. GFLOPS and parameters can comprehensively reflect the lightweight degree of the model, while FPS can reflect the real-time degree of the model.

[0038]

[0039] As can be seen from the table, adding c2f_star, ghostconv and lightweight decoupled detection head has a significant positive impact on the performance of the model. Specifically, in terms of detection accuracy: by gradually adding c2f_star, ghostconv and lightweight decoupled detection head, the mAP@0.5 of the model is increased from 0.824 of the basic model to 0.844, an increase of about 2 percentage points. This shows that the improvement measures effectively improve the detection accuracy of the model. In terms of model lightweighting: compared with the basic model, the model operation and parameter volume of the final version are almost halved. This shows that the computational requirements and storage requirements of the model are significantly reduced, and the model lightweighting effect is obvious. In terms of real-time performance: despite the model improvements, the final version of the model can still maintain a high FPS, which shows that the model can still maintain good real-time processing capabilities even when it is greatly lightweight.

[0040] In summary, by gradually introducing c2f_star, ghostconv and lightweight decoupled detection heads, we not only improved the detection accuracy of the model, but also significantly reduced the computational complexity and number of parameters of the model while maintaining high real-time processing capabilities. These improvements make the final version of the model more suitable for the actual application scenario of aluminum defect detection.

[0041] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the present invention should be included in the present invention.

Claims

1. A lightweight aluminum material defect detection method suitable for industrial environments, characterized by: Includes steps: Step 1: Obtain relevant labeled defect images from the aluminum defect dataset. The images contain defective parts and non-defective parts. Step 2: Perform data enhancement on the image; Step 3: Build a lightweight aluminum defect detection network, which includes a lightweight backbone network, a lightweight multi-scale feature fusion network, and a lightweight decoupling detection head; The lightweight backbone network is composed of a slice convolution module, four lightweight detection modules and a spatial pyramid pooling module in series; the lightweight detection module is composed of a lightweight convolution ghostconv and three lightweight feature processing modules c2f-star in series; Step 4: Use the aluminum defect dataset to train a lightweight aluminum defect detection network to obtain a lightweight aluminum defect detection model suitable for industrial environments; Step 5: Use the lightweight aluminum material defect detection model obtained in step 4 to detect the produced aluminum materials; Step 6: Identify possible defects in the produced aluminum.

2. The method for detecting defects in lightweight aluminum materials suitable for industrial environments according to claim 1, characterized in that: The lightweight convolution ghostconv described in step three is implemented by the following method: first, ordinary convolution is used to reduce the number of channels to half of the original number, and the obtained feature map is called the intrinsic feature map. Then, deep convolution is used to extract features for each channel of the intrinsic feature map. The extracted feature map is called the ghost feature map. Finally, the intrinsic feature map and the ghost feature map are added to obtain the output feature map.

3. The method for detecting defects in lightweight aluminum materials suitable for industrial environments according to claim 1, characterized in that: The lightweight feature processing module c2f-star described in step 3 is implemented by the following method: first, the number of output channels is multiplied by the expansion rate to obtain the number of hidden layer channels, and then the number of channels of the input feature map is transformed into twice the number of hidden layer channels through 1×1 convolution, and then the channels of the feature map are evenly divided into two parts, one of which is directly passed to the output, and the other part is input to the starblock module for processing; specifically, this part of the data first passes through a deep convolution layer with a convolution kernel size of 7, and then is divided into two paths, which are respectively increased in dimension through two fully connected layers f1 and f2. The output of f1 is activated using the ReLU6 activation function, and the output result of the ReLU6 activation is multiplied element by element with the output result of f2; the result of the multiplication is reduced to the dimension of the input through the fully connected layer g; then the output of the fully connected layer g is processed again using deep convolution, and the processing result is added to the input of the starblock module to obtain the final output.

4. The method for detecting defects in lightweight aluminum materials suitable for industrial environments according to claim 1, characterized in that: The lightweight multi-scale feature fusion network described in step three is composed of four feature fusion modules connected in series; the first two feature fusion modules in the four feature fusion modules are composed of an upsampling module, a feature map connection module and a lightweight feature processing module c2f-star connected in series; the last two feature fusion modules are composed of a convolution module, a feature map connection module and a lightweight feature processing module c2f-star connected in series; the feature map connection modules in the four feature fusion modules are respectively connected to the corresponding module input and the output of the 6th, 4th, 12th, and 9th layer models.

5. The method for detecting defects in lightweight aluminum materials suitable for industrial environments according to claim 1, characterized in that: The lightweight decoupled detection head described in step 3 is implemented by the following method: first, two serially connected group convolutions are used for feature extraction to achieve parameter sharing; then, the feature maps obtained by the group convolution are output to the 1×1 convolution layer for bounding box prediction and the 1×1 convolution layer for category prediction, respectively, and finally the bounding box loss and category loss are obtained for back propagation.

6. The method for detecting defects in lightweight aluminum materials suitable for industrial environments according to claim 1, characterized in that: The defects described in step 1 include dents, non-conductivity, scratches, orange peel, leakage, bruises, pits, convex powder, coating cracks and dirty spots.

7. The method for detecting defects in lightweight aluminum materials suitable for industrial environments according to claim 1, characterized in that: In the data enhancement processing of the image described in step 2, the method of data expansion of the image mainly includes: adding image noise, adjusting image brightness, image shearing, image rotation, image translation and mirroring, and finally expanding the data set to five times the original size.

Images