A Weld Image Enhancement Method for Motor Stator Core Based on Machine Vision

Abstract

The present invention proposes a weld seam image enhancement method for the stator core of an electric motor based on machine vision. The clarity of the weld seam image is improved through three modules: global feature extraction, local feature fusion, and detail enhancement. By using multi-scale convolution combined with an attention mechanism, the global semantic information of the weld seam image is extracted, enhancing the overall feature expression ability. Through convolution operations and a specific fusion strategy, the global features are effectively combined with local details to highlight the significant features of the weld seam area. A refinement network based on an edge attention mechanism is designed to focus on enhancing the edges and textures of the weld seam, improving the detail expressiveness of the image.

Description

Technical Field

[0001] The present invention belongs to the field of image enhancement, and particularly relates to a method for enhancing the weld image of the motor stator core based on machine vision. Background Art

[0002] The quality of the weld of the motor stator core directly affects the operating performance and service life of the motor. However, due to the complex welding environment, the weld surface is easily affected by noise, blurring, reflection, etc. Traditional image processing methods have limited effects in improving the quality of weld images. The contrast enhancement based on simple enhancement algorithms often fails to effectively highlight the features of the weld area. This limitation poses challenges to the accurate detection of weld defects.

[0003] Existing technologies mostly focus on single enhancement means and are difficult to balance the processing of global information and local details. For example, only increasing the contrast may lead to the amplification of background noise, and only focusing on local details may ignore the overall structure. Therefore, a comprehensive enhancement method that combines global features and local details is needed, and an attention mechanism is introduced to highlight the key areas of the weld, so as to better meet the requirements for improving the quality of complex weld images.

[0004] By proposing a method for enhancing the weld image of the motor stator core based on machine vision, the clarity and detail expressiveness of the weld image can be improved, providing high-quality input data for weld defect detection and quality assessment. The application of this method can not only improve the detection accuracy but also enhance the automation level of industrial production, which is of great significance for ensuring the quality of motor products and extending their service life. Summary of the Invention

[0005] The present invention provides a method for enhancing the weld image of the motor stator core based on machine vision, aiming to improve the clarity and detail expression ability of the weld image for more efficient defect detection and quality assessment. The method includes three modules: global feature extraction, local feature fusion, and detail enhancement. In the global feature extraction module, multi-scale convolution is combined with the attention mechanism to extract the global semantic information of the weld image and enhance the overall feature expression ability. In the local feature fusion module, the global feature and local details are effectively combined through convolution operations and specific fusion strategies to highlight the significant features of the weld area. In the detail enhancement module, a refinement network based on the edge attention mechanism is designed to focus on enhancing the edges and textures of the weld to improve the detail expressiveness of the image. This method can be widely applied to the field of detection and enhancement of industrial weld images, providing technical support for weld quality control.

[0006] The present invention aims to propose an optimized model for weld image enhancement and provides a method for enhancing the weld image of the motor stator core based on machine vision, including the following steps:

[0007] S1. Production of the weld seam image dataset. Select a high-resolution industrial camera, use a ring light source to reduce shadow and specular spot interference, and at the same time adapt to the reflection characteristics of weld seams of different materials. Collect images under an environment of uniform illumination and no strong light interference. Select two backgrounds, namely the factory site and the laboratory environment, and collect images of cracks in three different materials, namely steel, aluminum alloy, and stainless steel, to obtain the weld seam image dataset of the motor stator core;

[0008] S2. Preprocess the obtained original dataset, including performing rotation and flipping operations on the dataset to generate multi-angle and multi-direction images, obtaining the weld seam image dataset. The size of all images is unified to , and use the grayscale histogram and Gaussian filtering for preliminary image enhancement;

[0009] S3. Construct an attention module, including function calculation and dot product operation;

[0010] S4. Construct a global feature extraction module, including convolution,[[]] convolution operation,[[]] operation and activation function;

[0011] S5. Construct a detail enhancement module, including convolution operation,[[]] activation function, attention module,[[]] operation,[[]] operation and splicing operation;

[0012] S6. Construct a local feature fusion module, including convolution operation,[[]] convolution operation,[[]] operation,[[]] activation function and attention module;

[0013] S7. Construct a weld seam image enhancement model for the motor stator core based on machine vision, including a global feature extraction module, a detail enhancement module, a local feature fusion module and a splicing operation;

[0014] S8. Use the weld seam image dataset of the motor stator core to train a weld seam image enhancement model for the motor stator core based on machine vision.

[0015] Preferably, in S2, for each image in the original dataset, perform flipping and rotation at a random angle on it to obtain the weld seam image dataset of the motor stator core. The size of each image is unified to , which is convenient for model training. Use histogram equalization to process each image to obtain a crack image with enhanced contrast, and use Gaussian filtering to remove part of the interference in the image.

[0016] Preferably, in S3, an attention module is constructed. Denote the input image as , where represents the height, represents the width, represents the number of channels. Apply a function to all channels at each spatial position of the feature map to obtain a weight map with the same size as the original image but reflecting the relative importance of each channel . Multiply this weight map element-wise with the original feature map to get the weighted feature map , where represents the element-wise multiplication operation.

[0017] Preferably, using the attention module proposed in S3 enhances the significant features related to the weld seam, while also suppressing irrelevant or interfering channel information, which helps improve the performance of subsequent weld seam detection and quality assessment tasks.

[0018] Preferably, in S4, for the global feature extraction module, first denote the input weld image as , where represents the height of the feature map, represents the width of the feature map, represents the number of channels of the feature map. The feature map is processed by a convolutional neural unit module, and the unit module includes a convolution kernel, and activation function. After passing through this module, the feature map is obtained, where represents the convolution operation with a convolution kernel size of , represents of the convolution kernel, represents the bias term, represents operation, represents activation function. Next, the feature is further processed by a convolutional neural unit module, and this unit module includes a convolution kernel, and activation function. After passing through the unit module, the feature map is obtained, where represents the convolution operation with a convolution kernel size of , represents The convolution kernel, represents the bias term, represents operation, represents activation function, which multiplies the feature with the feature to perform a dot product operation to obtain the feature , where represents the dot product operation, and the feature is processed by the attention module constructed by S3 to obtain the feature , where represents the attention module, represents the convolution operation with a convolution kernel size of , represents the convolution kernel of, and multiplies again after the operation of the convolutional neural unit module to obtain , where represents the convolution operation with a convolution kernel size of , represents the convolution kernel of, represents the bias term, represents operation, represents activation function, which multiplies the obtained with to perform a dot product operation to obtain the feature map , where represents the dot product operation, is processed by the attention module constructed by S3 to obtain the feature , where represents the attention module, represents the convolution operation with a convolution kernel size of , represents the convolution kernel of, and finally multiplies after passing through the convolutional neural unit module to obtain the final global feature output , where represents the convolution operation with a convolution kernel size of , represents the convolution kernel of, represents the bias term, represents operation, represents activation function.

[0019] Preferably, the global feature extraction module proposed in S4 is used. Through multiple convolution and dot product operations, it combines low-level to high-level feature information. And through the attention module, it can effectively focus on the key regions in the image, enhancing the detailed performance of the welds in the image, thus improving the quality and usability of the image, helping to extract richer global features, and enhancing the enhancement effect of the weld image.

[0020] Preferably, in S5, for the detail enhancement module, the input feature map , where represents the height of the feature map, represents the width of the feature map, represents the number of channels of the feature map, The sum of all channel values at each spatial position is used to obtain a new feature map , where represents the -th channel of represents the value of the -th channel of . Next, the feature undergoes convolution operation and the activation function for linear transformation to obtain the feature , where represents the convolution operation with a convolution kernel size of represents the convolution kernel of , represents the activation function, After being processed by the attention module constructed in S3, the feature is obtained, where represents the attention module. The above steps are repeated again to obtain the feature . At the same time, the feature undergoes two consecutive max-pooling operations and self-attention mechanism operations to obtain , where represents the max-pooling operation with a pooling window size of , represents the self-attention function. Finally, the feature is concatenated with to obtain the final detail enhancement feature , where represents the concatenation operation.

[0021] Preferably, the detail enhancement module proposed by S5 is used. By splicing features of different scales and levels, the detail enhancement module can provide a richer feature representation, which helps to improve the performance of the model. Secondly, through convolution operations and attention mechanisms, the detail information in the weld image can be enhanced. Finally, through reasonable implementation and optimization, high computational efficiency can be maintained while ensuring performance.

[0022] Preferably, in S6, for the local feature fusion module, first, the global feature is concatenated with the detail enhancement feature to obtain , where represents the concatenation operation. Subsequently, the feature is further processed by the convolutional neural unit module, which includes a convolution kernel, and activation function. After passing through the unit module, the feature map is obtained, where represents the convolution operation with a convolution kernel size of , represents convolution kernel, represents the bias term, represents operation, represents activation function. Secondly, it is processed by the convolutional neural unit module, which includes a convolution kernel, and activation function. After passing through this module, the feature map is obtained, where represents the convolution operation with a convolution kernel size of , represents convolution kernel, represents the bias term, represents operation, represents activation function. Finally, the feature is processed by the attention module constructed by S3 to obtain the feature , where represents the attention module, represents convolution kernel.

[0023] Preferably, the local feature fusion module proposed by S6 is used. Through multi-stage feature processing, the effective combination of global semantic information and local detail features is achieved, and the key features of texture, edge, and defect in the weld image can be enhanced.

[0024] Preferably, in S7, for a weld seam image enhancement model of a motor stator core based on machine vision, the initially input weld seam image is , and is obtained through the global feature extraction module, where represents the global feature extraction module constructed in S4. is obtained through the detail enhancement module, where is represents the detail enhancement module constructed in S5. The global feature and the detail enhancement feature are concatenated to obtain , where represents the local feature fusion module constructed in S6. is concatenated with to obtain the final image enhancement optimization result , where represents the concatenation operation.

[0025] In summary, due to the adoption of this technical solution, compared with the prior art, the beneficial effects of the present invention are as follows: A weld seam image enhancement method for a motor stator core based on machine vision proposed by the present invention includes image preprocessing, an attention module, a global feature extraction module, a detail enhancement module, and a local feature fusion module. Image preprocessing expands the data set, performs contrast enhancement and Gaussian filtering on the image, improves the quality of the data set, and enhances the effect of model training. The attention module is proposed to enhance the significant features related to the weld seam and suppress the irrelevant or interfering channel information, improving the performance of subsequent weld seam detection and quality assessment tasks. The global feature extraction module is proposed to extract high-order semantic information and low-order detail information, effectively focus on the key regions in the image, enhance the detail performance of the target, thereby enhancing the quality of the image, and helping to extract richer global features. The detail enhancement module is proposed to concatenate image features of different scales, provide richer detail texture features, and combine the attention mechanism to enhance the features while improving the computational efficiency of the model and reducing the resource overhead. The local feature fusion module is proposed to comprehensively enhance the detail and semantic features of the overall weld seam image through the fusion of high-order features and low-order features in multiple stages. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flowchart of the steps of the weld seam image enhancement method for the motor stator core.

[0027] Figure 2 It is a structural diagram of the attention module.

[0028] Figure 3 It is a structural diagram of the global feature extraction module.

[0029] Figure 4 It is the structure diagram of the detail enhancement module.

[0030] Figure 5 It is the structure diagram of the local feature fusion module.

[0031] Figure 6 It is the structure diagram of the weld image enhancement model for the motor stator core based on machine vision.

[0032] Figure 7 It is a low-quality weld image of the motor stator core.

[0033] Figure 8 It is the optimized and enhanced weld image of the motor stator core. Specific implementation manner

[0034] The present invention aims to propose a weld image enhancement method for the motor stator core based on machine vision. The core lies in that data augmentation is performed on the data set to expand the diversity of the data, and at the same time improve the image quality. An attention module is introduced to strengthen the significant features related to the welds, while suppressing irrelevant or interfering channel information, significantly improving the model performance. A global feature extraction module is designed, which can capture high-order semantic information and low-order detail information, and effectively focus on the key areas in the image, enhancing the detail performance of the target and helping to extract richer global features. A detail enhancement module is proposed, which provides more abundant detail texture information by splicing image features of different scales. Combining with the attention mechanism, while enhancing the feature expression, it also improves the computational efficiency of the model. A local feature fusion module is proposed to achieve the deep fusion of high-order features and low-order features in multiple stages, comprehensively enhancing the overall details and semantic features of the weld image. Specifically, it includes the following steps, as Figure 1 shown.

[0035] S1. Organize the weld images of the motor stator core taken by a high-resolution industrial camera to obtain the initial weld image data set of the motor stator core. The images are collected in an environment with uniform illumination and no strong light interference, and are divided into two backgrounds: factory site and laboratory environment. Under each background, crack images of three different materials, namely steel, aluminum alloy, and stainless steel, are collected. There are 200 images for each material, and the initial data set has a total of 1200 images. The training set and the test set are divided according to a ratio of 4:1.

[0036] S2. Preprocess the obtained original data set, including performing rotation and flipping operations on the data set to generate multi-angle and multi-direction images, obtaining the weld image data set, and performing preliminary image enhancement using the grayscale histogram and Gaussian filtering.

[0037] Further, in S2, the obtained initial data set is preprocessed. First, each image is rotated by a random angle within the range of -170° to 170°, and 4 additional rotated images are obtained. Subsequently, each image is operated to obtain , and the size of the image is uniformly processed to size. Subsequently, histogram equalization is performed to obtain an image with enhanced contrast , and Gaussian filtering is performed to remove interference in some images, obtaining a preprocessed image . The preprocessed images form the data set for training, and the data set contains a total of 6000 images.

[0038] S3. Construct an attention module, which includes a function calculation and a dot product operation.

[0039] Further, in S3, for the attention module, its structure is as Figure 2 shown. For the input image , where 3 is the number of channels, and 224 and 224 are the width and height of the input image. For the input image , first perform a function calculation to obtain a channel weight feature vector , represents the weights of each pixel point on different channels. The higher the weight, the higher the importance of the corresponding pixel for enhancement. Multiply with for pixel-by-pixel multiplication of each channel to obtain a weighted calculated feature , where , , .

[0040] S4. Construct a global feature extraction module, which includes two convolutions, four convolution operations, four operations, and four activation functions.

[0041] Further, in S4, for the global feature extraction module, its structure is as Figure 3 shown. For the feature input to this module, , , , the feature vector first passes through a module that includes a convolution, and The neuron module of the activation function processes the preliminarily extracted features to obtain the extracted feature vector , , where , , , and then it enters the convolutional neuron unit containing convolution,[[]] and the activation function. This convolutional neuron unit calculates the feature vector , , , where , , . Then, and are dot-multiplied to obtain , , where , , . Then it undergoes one processing by the attention module to obtain , , where , , , thereby further extracting effective features. Then, it repeats the processing of the neuron module to obtain , has the same size as . Again, and are dot-multiplied to obtain , , where , , . Then, channel attention operation is performed on again to further enhance the weights of important channels and reduce the weights of non-critical channels, obtaining . After one convolution to adjust the number of channels, it passes through the neuron module for final feature extraction to obtain , , where , , which is the final global feature output.

[0042] S5. Build a detail enhancement module, which includes two convolution operations, two activation functions, two attention modules, two operations, two An operation and a splicing operation.

[0043] Furthermore, in S5, for the detail enhancement module, its structure is as Figure 4 shown. For the input of this module , first, the elements at the same position are added in the channel dimension to obtain the feature map after fusing features , , First, it passes through a convolution operation and an activation function for linear transformation to obtain the feature , whose channels are expanded, , then the attention module is used to process to extract effective information and obtain , , then the above operations are performed again to further extract features and obtain . At the same time, it passes through two consecutive max-pooling followed by a self-attention mechanism operation to obtain , , finally, the features and are spliced in the channel dimension to obtain , .

[0044] S6. Construct a local feature fusion module, which includes two convolution operations, one convolution operation, two operations, two activation functions and one attention module.

[0045] Furthermore, in S6, for the local feature fusion module, its structure is as Figure 5 shown. The input features of this module are the global features extracted by the global feature extraction module and the detail enhancement features output by the detail enhancement module , and First, they are spliced in the channel dimension to obtain , , then first, a convolutional neural unit module containing one convolution kernel, and an activation function is used for preliminary feature fusion to obtain the feature after preliminary fusion , , then the feature uses a convolutional neural unit module containing one Convolution kernel, and the convolution neural unit module of the activation function are further feature fused to obtain , , and then after one convolution fuses features and after one operation of the attention module, is obtained, , which is the enhanced image.

[0046] S7. Construct a weld image enhancement model of the motor stator core based on machine vision, including a global feature extraction module, a detail enhancement module, a local feature fusion module and a splicing operation.

[0047] Furthermore, in S7, for the weld image enhancement model of the motor stator core based on machine vision, its structure is as Figure 6 shown. First, the input weld image is , where , , , and through the global feature extraction module, is obtained, where represents the global feature extraction module constructed in S4, , where , , , , and through the detail enhancement module, is obtained, where represents the detail enhancement module constructed in S5, , the global feature and the detail enhancement feature are spliced to obtain , where represents the local feature fusion module constructed in S6, , and are spliced to obtain the final image enhancement optimization result , where represents the splicing operation, .

[0048] S8. Train the weld image enhancement optimization model of the motor stator core, and use the weld image dataset of the motor stator core to train the weld image enhancement optimization model of the motor stator core.

[0049] Furthermore, in S8, the operating system platform used for training the model is the Centos system, the language is Python 3.9.11, and the PyTorch deep learning framework is used for training. The hardware platform is the NVIDIA RTX 3090 with 24G of video memory. The dataset contains 6,000 crack images of three different materials, namely steel, aluminum alloy, and stainless steel, in two backgrounds: factory site and laboratory environment. The training set and test set are divided in a ratio of 4:1. During the training process, SGD is used as the optimizer, the initial learning rate is 0.001, and every 50 rounds of training, the learning rate decays to 0.1 of the original value, and a total of 500 rounds of training are performed.

[0050] Image processing of the motor stator core weld seam to obtain the motor stator core weld seam image that needs to be optimized and enhanced, as Figure 7 shown, Figure 7 shows a low-quality motor stator core weld seam image, which is input into the motor stator core weld seam image enhancement and optimization model to obtain the optimized and enhanced motor stator core weld seam image, as Figure 8 shown, Figure 8 shows the motor stator core weld seam image after being optimized and enhanced by the motor stator core weld seam image enhancement and optimization model.

[0051] The above is only the preferred implementation mode of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the creative concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A method for enhancing the weld image of a motor stator core based on machine vision, characterized in that: The following steps are involved: S1. Preparation of weld image dataset: Select high-resolution industrial camera to collect images in an environment with uniform lighting and no strong light interference. Select factory site and laboratory environment to collect images of cracks in three different materials: steel, aluminum alloy and stainless steel, and obtain the weld image dataset of the motor stator core. S2. Preprocess the original data set, including rotating and flipping the original images in the data set to obtain the weld image data set. All images are of the same size. , using grayscale histogram and Gaussian filtering for preliminary image enhancement, specifically, preprocessing the collected weld image dataset, first rotating and flipping each image in the dataset to obtain images in different orientations, thereby expanding the size of the dataset, and then processing the size of each image into size, and finally use histogram equalization and Gaussian filtering to process the image to enhance the contrast of the image and remove interference in the image to a certain extent, thereby improving the quality of the data set; S3, build attention module, including Function calculation and dot multiplication operation; S4, build a global feature extraction module, including convolution, Convolution operation, Operation and The activation function is specifically expressed as follows: the first input weld image is recorded as ,in represents the height of the feature map, represents the width of the feature map, The number of channels representing the feature map, the feature map Processed by convolutional neural unit module, the unit module includes a Convolution kernel, and Activation function, after passing through this module, the feature map is obtained ,in Indicates that the convolution kernel size is The convolution operation represents The convolution kernel, represents the bias term, express operate, express Activation function, next, feature It is further processed by a convolutional neural unit module, which consists of a Convolution kernel, and Activation function, after passing through the unit module, the feature map is obtained ,in Indicates that the convolution kernel size is The convolution operation, express The convolution kernel, represents the bias term, express operate, express Activation function, the feature With features Perform point multiplication to obtain features ,in Represents the dot product operation, feature After being processed by the attention module built by S3, the features are obtained ,in represents the attention module, Indicates that the convolution kernel size is The convolution operation, express The convolution kernel is After the convolutional neural unit module operation, we get ,in Indicates that the convolution kernel size is The convolution operation, express The convolution kernel, represents the bias term, express operate, express The activation function is obtained and Perform a dot multiplication operation to obtain the feature map ,in represents the dot product operation, The features are obtained through the attention module built by S3 ,in represents the attention module, Indicates that the convolution kernel size is The convolution operation, express The convolution kernel is finally The final global feature output is obtained through the convolutional neural unit module ,in Indicates that the convolution kernel size is The convolution operation, express The convolution kernel, represents the bias term, express operate, express Activation function; S5, build detail enhancement module, including Convolution operation, Activation function, attention module, operate, Operation and splicing operation, specifically expressed as, the input feature map ,in represents the height of the feature map, represents the width of the feature map, Indicates the number of channels of the feature map, All channel values ​​at each spatial position are added to obtain a new feature map ,in express No. channels, express No. The value of the channel, next, the feature go through Convolution operation and The activation function performs linear transformation to obtain features ,in Indicates that the convolution kernel size is The convolution operation, Expressed as The convolution kernel, represents the bias term, express Activation function, After being processed by the attention module built by S3, the features are obtained ,in Represents the attention module, repeat the above steps again to get the feature , while the characteristics After two consecutive maximum pooling operations and self-attention mechanism operations, we get ,in Indicates the pooling window size The maximum pooling operation, Represents the self-attention function. Finally, the feature and Stitching to get the final detail enhancement features ,in Represents a splicing operation; S6: Construct a local feature fusion module, including Convolution operation, Convolution operation, operate, Activation functions and attention modules; S7. Construct a weld image enhancement model of a motor stator core based on machine vision, including a global feature extraction module, a detail enhancement module, a local feature fusion module and a splicing operation; Among them, for the weld image enhancement model of the motor stator core based on machine vision, the weld image is first input as , obtained through the global feature extraction module ,in Represented as the global feature extraction module constructed in S4, Obtained through the detail enhancement module ,in Represented as the detail enhancement module constructed in S5, global features With detail enhancement features Splice to get ,in Represented as the local feature fusion module constructed in S6, and Stitching to get the final image enhancement optimization result ,in Represents a splicing operation; S8. Use the weld image dataset of the motor stator core to train a machine vision-based weld image enhancement model for the motor stator core.

2. The method for enhancing the weld seam image of a motor stator core based on machine vision according to claim 1, characterized in that: In S3, for the attention module, the input image is first recorded as ,in Indicates height, Indicates width, Indicates the number of channels, apply a Function on feature map All channels at each spatial position of the image are obtained to obtain a weight map with the same size as the original image but reflecting the relative importance of each channel , this weight map With the original feature map Multiply element by element to get the weighted feature map ,in Represents an element-wise multiplication operation.

3. The method for enhancing the weld seam image of a motor stator core based on machine vision according to claim 1, characterized in that: In S6, for the local feature fusion module, specifically expressed as: With detail enhancement features Splice to get ,in Represents the concatenation operation, and then the features It is further processed by a convolutional neural unit module, which consists of a Convolution kernel, and Activation function, after passing through the unit module, the feature map is obtained ,in Indicates that the convolution kernel size is The convolution operation, express The convolution kernel, represents the bias term, express operate, express activation function, and then processed by a convolutional neural unit module, which consists of a Convolution kernel, and Activation function, after passing through this module, the feature map is obtained ,in Indicates that the convolution kernel size is The convolution operation, express The convolution kernel, represents the bias term, express operate, express Activation function, final feature The features are processed by the attention module built by S3 ,in represents the attention module, express The convolution kernel.

4. The method for enhancing the weld seam image of a motor stator core based on machine vision according to claim 1, characterized in that: In S1, a ring light source is used to reduce shadow and spot interference, while adapting to the reflective characteristics of welds of different materials. Crack images of three different materials, steel, aluminum alloy and stainless steel, are collected. The background includes the factory site and laboratory environment, and the weld image dataset of the motor stator core is obtained.

Images