Furniture board edge sealing quality detection method and system based on machine vision

Abstract

The invention provides a furniture board edge sealing quality detection method and system based on machine vision. The method comprises the following steps: inputting a furniture board edge sealing image to be detected; the furniture board edge sealing image is preprocessed; performing quality identification by adopting a plate edge sealing quality identification model with an MLSF attention mechanism network to obtain a quality identification result; and the MLSF attention mechanism network linearly transforms the multi-scale features into vectors, performs multi-scale feature shunt convolution and spatial information enhancement to obtain attention output features of different scales, and then performs splicing and weighted summation on the attention output features of different scales to obtain output features. According to the method, the edge sealing quality of the furniture board can be automatically detected, the quality inspection efficiency in the production process is improved, and the labor cost is remarkably reduced; and by comprehensively considering the features of different scales, the recognition capability of the model on complex patterns is enhanced, and the detection precision and accuracy can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of visual inspection, and more specifically, to a method for inspecting the edge quality of furniture panels based on machine vision. Background Art

[0002] Nowadays, furniture manufacturing has shifted from rigid manufacturing to flexible manufacturing, and the characteristics of flexible manufacturing are small batch and multi-variety non-standard manufacturing; therefore, the quality inspection method of rigid manufacturing products is difficult to apply to the quality inspection process of flexible manufacturing products.

[0003] In the field of furniture manufacturing, although many production processes have been mechanized, such as automated board conveying, edge banding and board delivery, this automated process still has its limitations. Due to the lack of human intervention, problems such as misprocessing, edge bursting or incomplete edge banding that may occur during the production process become difficult to control. In the quality inspection link, the traditional manual visual inspection method is not only inefficient but also prone to errors, while sampling inspection may lead to missed inspections. In addition, this method also has problems in the consumption of human resources, which not only increases costs but also limits the improvement of production efficiency.

[0004] In contrast, machine vision inspection technology, with its significant advantages, provides an effective supplementary solution for quality inspection in the furniture manufacturing industry. However, due to the large number of colors, complex patterns, and diverse patterns of furniture veneer edge banding, traditional machine vision inspection algorithms face challenges in edge banding quality recognition precision and accuracy. Therefore, the current machine vision inspection technology for furniture board edge banding still needs further research and improvement.

[0005] This technology has become the preferred method for plate defect detection due to its cost-effectiveness, high accuracy, wide applicability and flexibility. Machine vision inspection can not only significantly improve the efficiency of quality inspection and reduce the risk of false detection and missed detection, but also greatly reduce the reliance on manual labor, thereby reducing labor costs. Through the application of this technology, furniture manufacturers can optimize production processes, improve product quality, and ultimately achieve significant benefits. Summary of the invention

[0006] In order to overcome the shortcomings and deficiencies in the prior art, the purpose of the present invention is to provide a method and system for detecting the edge banding quality of furniture panels based on machine vision; the method can realize automatic detection of the edge banding quality of furniture panels, improve the quality inspection efficiency in the production process, and significantly reduce labor costs; and by comprehensively considering the features of different scales, the model's recognition ability of complex patterns is enhanced, thereby improving the precision and accuracy of detection.

[0007] In order to achieve the above-mentioned object, the present invention is implemented by the following technical scheme: a furniture board edge banding quality detection method based on machine vision, inputting the furniture board edge banding image to be detected; preprocessing the furniture board edge banding image; using a board edge banding quality recognition model with an MLSF attention mechanism network to perform quality recognition on the preprocessed furniture board edge banding image, and obtaining a quality recognition result; The MLSF attention mechanism network obtains K, V, and Q vectors through linear transformation of multi-scale features, and performs multi-scale feature diversion convolution and spatial information enhancement to obtain attention output features of different scales. The attention output features of different scales are then concatenated and weighted summed to obtain the output features of the MLSF attention mechanism network.

[0008] Preferably, the MLSF attention mechanism network includes: a KVQ transformation module, a feature diversion convolution module and a spatial information enhancement module; The KVQ transformation module is used to obtain K, V, and Q vectors by linear transformation of multi-scale features respectively; The feature splitting convolution module is used to obtain enhanced multi-scale extraction features by performing dynamic convolution, multi-scale feature extraction, multiple activation functions and feature fusion on the K, V, Q vectors output by the KVQ transformation module; The spatial information enhancement module is used to use the features output by the feature diversion convolution module as input, perform GSConv processing to generate a feature map, map it to the spatial attention mechanism to adjust the contribution of different spatial positions to enhance the features of the key spatial area, use the ReLu activation function to increase the nonlinear expression ability, and then add the input features of the spatial information enhancement module to form a residual connection to obtain a spatial enhancement feature map; The outputs of multiple spatial information enhancement modules are weighted summed and concatenated to obtain the output features of the MLSF attention mechanism network.

[0009] Preferably, in the MLSF attention mechanism network, there are two feature diversion convolution modules; there are three spatial information enhancement modules, namely, a first spatial information enhancement module, a second spatial information enhancement module and a third spatial information enhancement module; The input of the MLSF attention mechanism network is connected to the input of the KVQ transformation module; the first and second output ends of the KVQ transformation module are connected one-to-one with the first spatial information enhancement module and the second spatial information enhancement module through two feature diversion convolution modules respectively; the third output end of the KVQ transformation module is directly connected to the third spatial information enhancement module; the output features of the second spatial information enhancement module are dot-multiplied with the output features of the first spatial information enhancement module and the second spatial information enhancement module respectively to obtain two dot-product features respectively; the two dot-product features are added to the input of the MLSF attention mechanism network respectively, and then spliced ​​to obtain the output features of the MLSF attention mechanism network.

[0010] Preferably, the feature splitting convolution module comprises: a convolution layer 1, three dynamic convolution layers of different scales, three different activation functions, a feature concatenation layer and a convolution layer 2; The input of the feature diversion convolution module is connected to the convolution layer 1, which passes through three dynamic convolution layers of different scales and is connected to different activation functions. After the activation function, they are connected to the feature concatenation layer together. Then, they are connected to the output of the feature diversion convolution module through the convolution layer 2.

[0011] Preferably, the three different activation functions are: ReLu, Sigmoid and Tanh.

[0012] Preferably, the spatial information enhancement module includes a GSConv and a spatial attention mechanism unit connected to each other; the GSConv is connected to the input of the spatial information enhancement module; the output of the spatial attention mechanism unit is connected to the input residual of the spatial information enhancement module; and then the output of the spatial information enhancement module is connected.

[0013] Preferably, the board edge banding quality recognition model refers to a model in which an MLSF attention mechanism network is added to the YOLOv5 model; the MLSF attention mechanism network is arranged between the neck network output and the head network input of the YOLOv5 model.

[0014] Preferably, after obtaining the quality identification result, if the quality identification result is that there is a defect, the quality identification result is recorded in the database and notified for processing.

[0015] Preferably, the furniture board edge banding image to be detected refers to: using multiple cameras to capture images of the furniture board edge banding position; integrating the images of multiple cameras into a fused image through image fusion technology; and using the fused image as the furniture board edge banding image to be detected.

[0016] A system comprises a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the machine vision-based furniture board edge banding quality detection method is implemented.

[0017] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. The present invention combines the diversity of the feature diversion convolution module and the focusing ability of the spatial information enhancement module, and can capture and focus on important feature information at multiple scales; the MLSF attention mechanism network can more effectively capture the detailed features of the edge sealing of furniture panels, and can achieve more refined and accurate detection results in terms of both color distinction and pattern recognition, and can detect various edge sealing defects more accurately, quickly and comprehensively; 2. In the present invention, the feature split convolution module significantly enhances the feature extraction capability of the model through dynamic convolution, multi-scale feature extraction, multiple activation functions and feature fusion; compared with the ordinary convolution module, the feature split convolution module performs better in processing complex scenes, adapting to input changes and extracting rich features. By splicing and fusing the features of multiple branches, the complexity and expression of the features are increased, and the discrimination ability of the model is improved; 3. In the spatial information enhancement module of the present invention, the channels in each group in GSConv are independently convolved to extract features, which helps to reduce the number of parameters and the amount of calculation while retaining the diversity of features; then the feature map generated by GSConv is mapped to the spatial attention mechanism to adjust the contribution of different spatial positions to enhance the features of the key spatial area, and the ReLu activation function is used to enhance the feature map of the spatial attention mechanism to increase the nonlinear expression ability of the network; then the features enhanced by the ReLu nonlinearity are added to the input features of the spatial information enhancement module to form a residual connection to obtain a spatial enhanced feature map; the network learns a deeper feature representation and alleviates the gradient vanishing problem; the spatial information enhancement module can effectively enhance the network's perception of spatial information and improve the performance of the model in processing visual tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a structural diagram of the MLSF attention mechanism network of the furniture board edge banding quality detection method based on machine vision of the present invention; Figure 2 It is a structural diagram of a feature diversion convolution module of a furniture board edge sealing quality detection method based on machine vision of the present invention; Figure 3 It is a structural diagram of a spatial information enhancement module of a furniture board edge sealing quality detection method based on machine vision of the present invention; Figure 4It is a structural diagram of a furniture board edge sealing quality detection method based on machine vision and a board edge sealing quality recognition model of the present invention; Figure 5 It is a flow chart of the edge sealing quality detection method of furniture board based on machine vision of the present invention; Figure 6 It is a schematic diagram of the structure of the detection platform of the furniture board edge sealing quality detection method based on machine vision of the present invention; Figure 7 It is a schematic diagram of image fusion of the furniture board edge banding quality detection method based on machine vision of the present invention. DETAILED DESCRIPTION

[0019] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0020] Embodiment 1 The present embodiment provides a furniture board edge banding quality detection method based on machine vision, which includes inputting an edge banding image of a furniture board to be detected; preprocessing the edge banding image of the furniture board; and using a board edge banding quality recognition model with an MLSF attention mechanism network to perform quality recognition on the preprocessed edge banding image of the furniture board to obtain a quality recognition result.

[0021] Since furniture board edge banding has many colors, complex patterns, and diverse patterns, traditional target detection algorithms face challenges in recognition precision and accuracy. In order to cope with the challenge of complex patterns to detection precision, this paper proposes a board edge banding quality recognition model with an MLSF (Multi-scale Linear Space Fusion Attention) attention mechanism network, the purpose is to enhance the model's recognition ability for complex patterns by comprehensively considering features of different scales, thereby improving the precision and accuracy of detection. Through this mechanism, the detailed features of furniture board edge banding can be captured more effectively, and more refined and accurate detection results can be achieved in both color distinction and pattern recognition.

[0022] The MLSF attention mechanism network obtains K, V, and Q vectors through linear transformation of multi-scale features, and performs multi-scale feature diversion convolution and spatial information enhancement to obtain attention output features of different scales. The attention output features of different scales are then concatenated and weighted summed to obtain the output features of the MLSF attention mechanism network.

[0023] The MLSF attention mechanism network includes: KVQ transformation module, feature diversion convolution module and spatial information enhancement module, such as Figure 1 shown.

[0024] The KVQ transformation module is used to obtain K, V, and Q vectors through linear transformation of multi-scale features; The feature splitting convolution module is used to obtain enhanced multi-scale extraction features from the K, V, Q vectors output by the KVQ transformation module through dynamic convolution, multi-scale feature extraction, multiple activation functions and feature fusion.

[0025] like Figure 2 As shown, the feature splitting convolution module includes: convolution layer 1, three dynamic convolution layers of different scales, three different activation functions, feature concatenation layer and convolution layer 2; The input features of the feature diversion convolution module are convolved through the convolution layer one, and then dynamically convolved through three dynamic convolution layers of different scales to achieve multi-scale feature extraction. The extracted features of different scales are connected to different activation functions respectively; the features processed by the activation function are input to the feature splicing layer for feature fusion; then convolution is performed through the convolution layer two to obtain enhanced multi-scale extracted features as the output of the feature diversion convolution module.

[0026] The feature splitting convolution module significantly enhances the model's feature extraction capability through dynamic convolution, multi-scale feature extraction, multiple activation functions, and feature fusion. Compared with ordinary convolution modules, the feature splitting convolution module performs better in processing complex scenes, adapting to input changes, and extracting rich features. By splicing and fusing the features of multiple branches, the complexity and expression of the features are increased, and the model's discrimination ability is improved. These advantages enable it to achieve better results in furniture board edge banding quality inspection tasks.

[0027] The spatial information enhancement module is used to take the features output by the feature diversion convolution module as input, perform GSConv processing to generate feature maps, and perform convolution operations on the channels in each group in GSConv to extract features independently, which helps to reduce the number of parameters and the amount of calculation while retaining the diversity of features; then the feature map generated by GSConv is mapped to the spatial attention mechanism to adjust the contribution of different spatial positions to enhance the features of the key spatial area, and the ReLu activation function is used to increase the nonlinear expression ability of the network for the feature map enhanced by the spatial attention mechanism; then the features enhanced by the ReLu nonlinearity are added to the input features of the spatial information enhancement module to form a residual connection to obtain the spatial enhanced feature map. This enables the network to learn deeper feature representations and alleviate the problem of gradient disappearance. Finally, the enhanced feature map is output and passed to the next layer of the network. Through the above process, the spatial information enhancement module can effectively enhance the network's perception of spatial information and improve the performance of the model in processing visual tasks.

[0028] like Figure 3As shown, the spatial information enhancement module includes a connected GSConv and a spatial attention mechanism unit; the GSConv is connected to the input of the spatial information enhancement module; the output of the spatial attention mechanism unit is connected to the input residual of the spatial information enhancement module; and then the output of the spatial information enhancement module is connected.

[0029] In the MLSF attention mechanism network, there are two feature diversion convolution modules and three spatial information enhancement modules, namely the first spatial information enhancement module, the second spatial information enhancement module and the third spatial information enhancement module; The input of the MLSF attention mechanism network is connected to the input of the KVQ transformation module; the first and second output ends of the KVQ transformation module are connected one-to-one with the first spatial information enhancement module and the second spatial information enhancement module through two feature diversion convolution modules respectively; the third output end of the KVQ transformation module is directly connected to the third spatial information enhancement module; the output features of the second spatial information enhancement module are dot-multiplied with the output features of the first spatial information enhancement module and the second spatial information enhancement module respectively to obtain two dot-product features respectively; the two dot-product features are added to the input of the MLSF attention mechanism network respectively, and then spliced ​​to obtain the output features of the MLSF attention mechanism network.

[0030] The present invention combines the diversity of the feature diversion convolution module and the focusing ability of the spatial information enhancement module, and can capture and focus on important feature information at multiple scales. The application of the MLSF attention mechanism network can enable the furniture board edge banding quality inspection system to detect various edge banding defects more accurately, quickly and comprehensively. The application of this technology not only improves the efficiency and accuracy of inspection, but also provides a powerful tool for the furniture manufacturing industry to improve product quality and production efficiency.

[0031] In order to refine the detection results and enable the model to more accurately classify and locate edge defects, the plate edge quality recognition model of the present invention refers to a model that adds an MLSF attention mechanism network on the basis of the YOLOv5 model; the MLSF attention mechanism network is set between the neck network output and the head network input of the YOLOv5 model, such as Figure 4 shown.

[0032] The trained original YOLOv5 model and the board edge banding quality recognition model of the present invention were tested on furniture board edge banding image samples, and the test results are shown in Table 1.

[0033] Table 1 Comparison of test results between the original YOLOv5 model and the panel edge quality recognition model of the present invention It can be seen from Table 1 that the panel edge quality recognition model of the present invention is superior to the original YOLOv5 model in terms of P precision, R recall, F1 score and mAP50 indicators.

[0034] Furniture board edge banding quality inspection process, such as Figure 5 As shown, the following steps are included: First, the detection platform is initialized. Figure 6 As shown; a camera group 1 is constructed by using multiple cameras, and in this embodiment, there are four cameras; in the process of conveying furniture boards by the conveyor belt 4, the camera group 1 collects images of the edge sealing positions of the furniture boards; multiple cameras are respectively fixed in the fixtures and the camera positions are adjusted to ensure that the cameras are aimed at the object to be measured at the most ideal angle and position, so as to capture the clearest image; by adjusting the emission angle of the light source 2, the image acquisition conditions are optimized to ensure that the surface features and detail information of the object to be measured can be most effectively displayed when the image is acquired.

[0035] The barcode scanner 3 is used to identify the QR code attached to the surface of the furniture board, and the corresponding unique identification code is retrieved from the database to obtain the board database information; according to the board database information, the brightness and color of the light source 2 are dynamically adjusted to adapt to the changing detection targets and complex lighting environments. This intelligent light source adjustment ensures the controllability of image acquisition parameters and improves the adaptability and flexibility in diversified field applications.

[0036] The images from multiple cameras are integrated into a fused image through image fusion technology, such as Figure 7 As shown in the figure, the fused image is used as the edge banding image of the furniture board to be inspected. This integration process not only optimizes the comprehensive utilization of resources, but also significantly improves the quality and availability of image data.

[0037] Preprocess the edge banding images of furniture panels to ensure optimal image quality. Preprocessing includes image light correction to balance lighting conditions, camera calibration to calibrate lens distortion, and image denoising to remove noise interference in the image. Together, these steps improve the clarity and accuracy of the image, laying a solid foundation for subsequent analysis.

[0038] The board edge banding quality recognition model with MLSF attention mechanism network is used to perform quality recognition on the preprocessed furniture board edge banding images and obtain the quality recognition results.

[0039] After obtaining the quality identification result, if the quality identification result is defective, the quality identification result will be recorded in the database and notified for processing.

[0040] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A method for detecting edge quality of furniture panels based on machine vision, characterized in that: Input the furniture board edge banding image to be detected; preprocess the furniture board edge banding image; use the board edge banding quality recognition model with MLSF attention mechanism network to perform quality recognition on the preprocessed furniture board edge banding image to obtain the quality recognition result; The MLSF attention mechanism network obtains K, V, and Q vectors through linear transformation of multi-scale features, and performs multi-scale feature diversion convolution and spatial information enhancement to obtain attention output features of different scales. The attention output features of different scales are then concatenated and weighted summed to obtain the output features of the MLSF attention mechanism network.

2. The method for detecting edge quality of furniture panels based on machine vision according to claim 1, characterized in that: The MLSF attention mechanism network includes: a KVQ transformation module, a feature diversion convolution module and a spatial information enhancement module; The KVQ transformation module is used to obtain K, V, and Q vectors by linear transformation of multi-scale features respectively; The feature splitting convolution module is used to obtain enhanced multi-scale extraction features by performing dynamic convolution, multi-scale feature extraction, multiple activation functions and feature fusion on the K, V, Q vectors output by the KVQ transformation module; The spatial information enhancement module is used to use the features output by the feature diversion convolution module as input, perform GSConv processing to generate a feature map, map it to the spatial attention mechanism to adjust the contribution of different spatial positions to enhance the features of the key spatial area, use the ReLu activation function to increase the nonlinear expression ability, and then add the input features of the spatial information enhancement module to form a residual connection to obtain a spatial enhancement feature map; The outputs of multiple spatial information enhancement modules are weighted summed and concatenated to obtain the output features of the MLSF attention mechanism network.

3. The method for detecting edge quality of furniture panels based on machine vision according to claim 2, characterized in that: In the MLSF attention mechanism network, there are two feature diversion convolution modules; there are three spatial information enhancement modules, namely, the first spatial information enhancement module, the second spatial information enhancement module and the third spatial information enhancement module; The input of the MLSF attention mechanism network is connected to the input of the KVQ transformation module; the first and second output ends of the KVQ transformation module are connected one-to-one with the first spatial information enhancement module and the second spatial information enhancement module through two feature diversion convolution modules respectively; the third output end of the KVQ transformation module is directly connected to the third spatial information enhancement module; the output features of the second spatial information enhancement module are dot-multiplied with the output features of the first spatial information enhancement module and the second spatial information enhancement module respectively to obtain two dot-product features respectively; the two dot-product features are added to the input of the MLSF attention mechanism network respectively, and then spliced ​​to obtain the output features of the MLSF attention mechanism network.

4. The method for detecting edge quality of furniture panels based on machine vision according to claim 2, characterized in that: The feature splitting convolution module includes: a convolution layer 1, three dynamic convolution layers of different scales, three different activation functions, a feature splicing layer and a convolution layer 2; The input of the feature diversion convolution module is connected to the convolution layer 1, which passes through three dynamic convolution layers of different scales and is connected to different activation functions. After the activation function, they are connected to the feature concatenation layer together. Then, they are connected to the output of the feature diversion convolution module through the convolution layer 2.

5. The method for detecting edge quality of furniture panels based on machine vision according to claim 4, characterized in that: The three different activation functions are: ReLu, Sigmoid and Tanh.

6. The method for detecting edge quality of furniture panels based on machine vision according to claim 2, characterized in that: The spatial information enhancement module includes connected GSConv and spatial attention mechanism units; The GSConv is connected to the input of the spatial information enhancement module; The output of the spatial attention mechanism unit is connected to the input residual of the spatial information enhancement module; then the output of the spatial information enhancement module is connected.

7. The method for detecting edge quality of furniture panels based on machine vision according to claim 1, characterized in that: The board edge banding quality recognition model refers to a model in which an MLSF attention mechanism network is added to the YOLOv5 model; the MLSF attention mechanism network is set between the neck network output and the head network input of the YOLOv5 model.

8. The method for detecting edge quality of furniture panels based on machine vision according to claim 1, characterized in that: After obtaining the quality identification result, if the quality identification result is defective, the quality identification result will be recorded in the database and notified for processing.

9. The method for detecting edge quality of furniture panels based on machine vision according to claim 1, characterized in that: The furniture board edge banding image to be detected refers to: using multiple cameras to collect images of the furniture board edge banding position; integrating the images of multiple cameras into a fused image through image fusion technology; and using the fused image as the furniture board edge banding image to be detected.

10. A system comprising a processor and a memory for storing a program executable by the processor, characterized in that: When the processor executes the program stored in the memory, the machine vision-based furniture board edge banding quality detection method described in any one of claims 1 to 9 is implemented.

Images