Method and system for detecting fine short circuit of circuit board based on machine vision

Abstract

The application discloses a kind of line board subtle short circuit detection method and system based on machine vision, comprising:A1: collecting line board original optical image data, line board image illumination intensity data, and carrying out cleaning;A2: by microdefect enhancement network, calculate defect enhancement feature, calculate illumination perception attention weight, dynamic segmentation threshold after illumination calibration, extract optimized defect candidate feature;A3: extract multi-dimensional initial defect feature, coupled multi-dimensional defect feature, then extract line board subtle short circuit defect feature;A4: extract weight optimized fusion feature, calculate calibrated defect confidence, and obtain effective defect region mask by threshold screening;A5: effective region screening is carried out, and the boundary box vertex coordinates of defect region, the actual physical size of defect are determined, to obtain line board subtle short circuit defect detection result.The application can solve the problem that traditional method is difficult to accurately identify subtle short circuit defect.

Description

Technical Field

[0001] This invention relates to the field of circuit board testing technology, and in particular to a method and system for detecting minute short circuits in circuit boards based on machine vision. Background Technology

[0002] As a core component of electronic devices, the production quality of circuit boards directly determines the operational stability and lifespan of electronic devices. Minor short circuit defects are common hidden faults in the circuit board production process. Due to their small size and the fact that they are often obscured by background information such as the texture and solder joints on the circuit board surface, they have become a key issue affecting the quality of circuit boards. Correspondingly, the micro short circuit detection technology for circuit boards is an important part of the quality control system in the electronics manufacturing industry.

[0003] Currently, the traditional methods used in the industry for short-circuit detection on printed circuit boards mainly include manual visual inspection and traditional machine vision inspection. Manual visual inspection relies on inspectors to observe and identify the surface of the printed circuit board with their naked eyes and experience. This method is not only labor-intensive and inefficient, but also easily affected by factors such as the fatigue level of the inspectors and differences in subjective judgment. It is difficult to accurately identify minute short-circuit defects with small size, resulting in a high rate of missed detection. Traditional machine vision inspection methods are mostly based on traditional image processing algorithms such as threshold segmentation, edge extraction, and morphological processing. They identify defects by setting fixed feature thresholds. However, they are not adaptable to changes in the lighting environment of the inspection environment and interference from stains on the surface of the printed circuit board. They are also difficult to effectively extract the differentiated features of minute short-circuit defects, and thus cannot meet the needs of accurate detection of minute short-circuit defects.

[0004] With the rapid development of deep learning technology in the field of image recognition, defect detection methods based on convolutional neural networks are gradually being applied to the field of circuit board inspection. Existing technologies typically involve collecting circuit board images to construct a defect dataset, training a convolutional neural network to enable it to recognize defects, and achieving automated detection. However, most of these existing solutions focus on optimizing models for obvious or large-sized defects on circuit boards, lacking targeted design for subtle short-circuit defects. They also fail to fully integrate efficient image enhancement techniques to strengthen the contrast between subtle short-circuit defect areas and the background, resulting in subtle defect features being easily masked by background information, making it difficult for the model to effectively capture them. At the same time, the model training process does not focus on mining subtle defect features, leading to insufficient accuracy in recognizing subtle short-circuit defects and a still high false negative rate, failing to meet the actual needs of the electronics manufacturing industry for high-precision circuit board inspection. Summary of the Invention

[0005] In view of this, the present invention aims to provide a method and system for detecting minute short circuits on circuit boards based on machine vision, so as to solve the problem that traditional methods are difficult to accurately identify minute short circuit defects.

[0006] A machine vision-based method for detecting minute short circuits on circuit boards includes: A1: Collect the original optical image data and illumination intensity data of the circuit board, and clean them to obtain the cleaned optical image and illumination intensity data of the circuit board. A2: Based on the optical image of the cleaned circuit board, a micro-defect enhancement network is used to calculate defect enhancement features, thereby enhancing the contrast between the region of minor short-circuit defects and the background to highlight the defect features and solving the problem that the features of minor short-circuit defects in existing solutions are easily masked by background information. Combined with the light intensity data of the cleaned circuit board, the light perception attention weight and the dynamic segmentation threshold after light calibration are calculated. The defect region segmentation and calibration are performed through a defect region calibration network to obtain the original defect candidate region features, and then the optimized defect candidate features are extracted. A3: Based on the optimized defect candidate features and the cleaned optical image of the circuit board, multidimensional initial defect features are extracted; combined with the illumination perception attention weight, coupled multidimensional defect features are extracted; then, through the texture suppression network, the pure defect features after interference removal are extracted, and finally, the minor short-circuit defect features of the circuit board are extracted. A4: Based on the optimized defect candidate features and the circuit board minor short-circuit defect features, extract the weighted fusion features, calculate the calibrated defect confidence, and obtain the effective defect area mask through threshold screening. A5: Valid regions are filtered by using a mask of valid defect regions, and the coordinates of the vertices of the bounding box of the defect region are determined. Then, the actual physical size of the defect is calculated to obtain the detection results of minor short-circuit defects on the circuit board.

[0007] Furthermore, step A1 also includes: The original optical image data of the circuit board is acquired by using a pixel industrial CCD camera with a ring-shaped LED uniform light source. The data type is two-dimensional grayscale pixel data. The optical image includes the front surface image and the back surface image of the circuit board. The original optical image data of the circuit board is preprocessed by Gaussian filtering noise reduction method to obtain the cleaned optical image of the circuit board. The illumination intensity data of the circuit board image is collected by an ambient light sensor. The data type is numerical data, which includes the ambient light intensity value at the time of collection. Then, the illumination intensity data of the circuit board after cleaning is obtained by removing constant values ​​of three times standard difference and smoothing by sliding window mean.

[0008] Furthermore, the micro-defect enhancement network in step A2 also includes: The cleaned circuit board optical image is processed by multi-scale convolution through convolutional layers of size 1×1, 3×3, and 5×5, respectively. The results are then added element-wise to obtain the multi-scale convolutional fusion feature. After processing the cleaned circuit board optical image through a multilayer perceptron, the result calculated by the Sigmoid function is combined with the multi-scale convolutional fusion feature by a Hadamard product operation. Finally, this result is added to the result of the cleaned circuit board optical image after processing by the residual connection layer to obtain the defect enhancement feature.

[0009] Furthermore, the defect area calibration network in step A2 also includes: The defect enhancement features are calculated using a graph convolutional network and the ReLU function, and then a Boolean matrix generation operation is performed with the dynamic segmentation threshold after illumination calibration. The defect binary mask is then obtained through Hadamard product operation. The defect enhancement features and the defect binary mask are then subjected to Hadamard product operation to complete the defect region segmentation and calibration, and the original defect candidate region features are obtained.

[0010] Furthermore, step A2 also includes: A21: Based on the optical image of the cleaned circuit board, the defect enhancement features are calculated using a micro-defect enhancement network. The calculation method of the micro-defect enhancement network is as follows: in, For multi-scale convolutional fusion features, A 1×1 convolutional layer, This is an optical image of the circuit board after cleaning. To add element by element, It is a 3×3 convolutional layer. It is a 5×5 convolutional layer. For defect enhancement features, For Hadama accumulation, For the Sigmoid function, It is a multilayer perceptron. For residual connection layer; A22: Based on the defect enhancement features and the illumination intensity data of the cleaned circuit board, calculate the illumination perception attention weight and the dynamic segmentation threshold after illumination calibration. The calculation method is as follows: in, For light perception attention weights, For splicing operations, For dimensional expansion operations, The data shows the light intensity of the circuit board after cleaning. This is the dynamic segmentation threshold after illumination calibration. For global average pooling; A23: Based on the defect enhancement features, illumination-aware attention weights, and the dynamic segmentation threshold after illumination calibration, defect region segmentation and calibration are performed using a defect region calibration network to obtain the original defect candidate region features. The calculation method of the defect region calibration network is as follows: in, For defect binary mask, For ReLU function, For graph convolutional networks, For Boolean matrix generation operations, Candidate features for original defects; A24: Based on the original candidate defect features, extract the optimized candidate defect features. The calculation method is as follows: in, Weights are used to verify the validity of candidate regions. For gating attention mechanism, These are the optimized candidate features for defects.

[0011] It should be further explained that the size of the micro short-circuit defects on the circuit board is very small, and the grayscale difference between the defect area and the background information such as the texture of the circuit board substrate and surface residues is very small. At the same time, the optical image of the circuit board after cleaning is very prone to the loss of effective defect details in the subsequent feature processing, which causes the micro short-circuit defect features to be submerged by the background information, making it difficult to extract accurately and greatly hindering the subsequent segmentation and calibration of the defect area. This invention constructs a micro-defect enhancement network, achieving multiple technical effects through the synergistic cooperation of multi-scale convolution processing, feature activation operations, and residual connections. First, multi-scale convolution processing is performed on the cleaned circuit board optical image using convolutional layers of different sizes to capture local subtle defect features, mid-level texture features, and global background features. These features are then element-wise added to obtain multi-scale convolutional fusion features, achieving comprehensive coverage of features across different dimensions and avoiding the problem of missing key defect information in single-scale convolution. Second, after processing the cleaned circuit board optical image using a multilayer perceptron, the network is then processed by Si... The gmoid function generates activation features, which are then combined with multi-scale convolutional features using a Hadamard product to dynamically amplify the feature differences between the defect region and the background region, making minute short-circuit defect features more prominent and effectively solving the problem of low defect feature recognition. Finally, the optical image of the cleaned circuit board is processed through a residual connection layer, and the processing result is added to the aforementioned calculation result. This maximizes the preservation of defect details in the original image, avoids the loss of effective information caused by multi-layer operations during feature enhancement, ensures the integrity of defect features, and significantly improves the quality of defect enhancement features. Existing technologies for enhancing circuit board image defects mostly employ single-scale convolution processing or simple image brightening and contrast adjustment. These methods either only capture single-dimensional features, failing to extract features from both subtle defects and background information, or only achieve global-level feature adjustment, unable to differentiate enhancement for subtle short-circuit defects. This results in blurred defect features and a high risk of losing key details during processing, while lacking an effective detail preservation mechanism, further reducing the usability of the enhanced features. In contrast, the micro-defect enhancement network of this invention innovatively uses multi-scale convolutional fusion to achieve comprehensive feature extraction, highlights defect features through feature activation operations, and fully preserves defect details through residual connections. This not only overcomes the limitations of existing single-processing methods but also solves the core problems of incomplete defect feature extraction, low recognition, and severe detail loss in existing technologies. It can output higher-quality defect enhancement features, laying a solid foundation for subsequent illumination-sensing attention weight calculation, dynamic segmentation threshold generation, and defect region calibration.

[0012] In addition, the optical image of the cleaned circuit board is easily affected by the fluctuation of the lighting environment. The gray difference between the fine short-circuit defect and the substrate texture is very small. Changes in the light intensity will further amplify this feature confusion, causing the gray features of the defect area to be distorted. At the same time, the traditional segmentation threshold cannot adapt to the image features under different lighting conditions, which can easily cause the defect area to be missed or the normal area to be misjudged as a false defect. To address the aforementioned issues, this invention designs a light-sensing attention weight and a dynamically segmented threshold after light calibration, achieving a dual effect of light adaptation and precise threshold generation: First, considering the dimensionality difference between the light intensity data and defect enhancement features of the cleaned circuit board, the dimension of the light intensity data is adjusted through a dimension expansion operation, enabling it to be concatenated with the defect enhancement features. This allows subsequent processing to simultaneously consider both the defect's inherent characteristics and the influence of ambient light. Second, the concatenated fused data undergoes convolution processing, followed by Sigmoid function calculation to obtain the light-sensing attention weight, accurately capturing the mapping relationship between light intensity and defect features, and enabling the segmentation of areas with abnormal light. Key features are highlighted to provide a targeted basis for subsequent threshold calibration. Finally, the enhanced features of defects are processed by global average pooling to obtain feature data reflecting the overall gray level of the defects. This data is then fused with the illumination-aware attention weights to generate a dynamic segmentation threshold after illumination calibration. This threshold is adjusted differently according to the illumination conditions of different regions. Reasonable segmentation sensitivity is maintained in uniformly illuminated areas, while feature misjudgment is avoided in areas with abnormal illumination through weight correction. The entire process, from illumination data adaptation and attention weight generation to dynamic threshold calibration, not only solves the problem of fusion between illumination data and image features but also achieves dynamic adaptation of the segmentation threshold, effectively improving the accuracy of defect region segmentation. Existing technologies for segmenting circuit board images mostly employ fixed segmentation thresholds or generate a single threshold based solely on the image's grayscale features. These methods generally ignore the impact of illumination intensity on image features, failing to adapt to illumination fluctuations in different detection scenarios or to differentiate the illumination differences in different regions of the image. Even when considering illumination factors, some technologies simply perform global illumination compensation without deeply integrating illumination data with defect features, resulting in a lack of specificity in the segmentation threshold and persistent issues such as missed defects and false defect detections. In contrast, this invention achieves deep fusion of illumination data and defect features by generating illumination-aware attention weights and a dynamically segmented threshold after illumination calibration. The dynamic segmentation threshold can be precisely adjusted according to illumination conditions, overcoming the limitations of fixed thresholds in existing technologies and solving the core problem of defect feature confusion under illumination interference. This significantly improves the accuracy and robustness of defect region segmentation, providing a high-quality threshold basis for the subsequent defect region calibration network to generate a binary defect mask.

[0013] Finally, the enhanced defect features after feature enhancement still contain interference information such as substrate texture and residual light, and the boundary contours of minor short-circuit defects are blurred. At the same time, the feature boundaries between the defect area and the normal circuit area are difficult to define accurately. Traditional segmentation methods are very likely to misjudge normal circuits as defects or miss minor defects. The calibration process needs to take into account the integrity and purity of defect features. If not handled properly, it will lead to distortion of subsequent defect feature extraction and affect the accuracy of the final detection results. This invention constructs a defect region calibration network. Through the synergistic linkage of graph convolutional networks, ReLU functions, Boolean matrix generation operations, and defect binary mask calibration, a logical chain is formed from feature optimization and region selection to precise calibration. First, addressing the issue of insufficiently exploited spatial correlation features between defects and the background in defect enhancement features, a graph convolutional network is used to process the defect enhancement features, capturing spatial correlation information within the defect region and between the defect and the background, strengthening the overall feature expression of the defect region, and weakening isolated background interference features. Subsequently, ReLU function activation processing is applied to suppress invalid negative information in the features, highlighting the positive features. First, based on the dynamic segmentation threshold after illumination calibration, regions that meet the threshold requirements in the defect enhancement features are selected through Boolean matrix generation operations, locking the defect candidate range and providing a clear target for subsequent calibration. This Boolean matrix can respond to the dynamically generated threshold in the early stage and adapt to the defect distribution characteristics under different illumination conditions. Finally, the feature map processed by graph convolutional network and ReLU function is fused with the defect binary mask. The defect enhancement features are accurately selected and calibrated with the help of the defect binary mask, eliminating the interference areas outside the mask and retaining the effective defect features within the mask to obtain the original defect candidate region features. Existing technologies for circuit board defect region segmentation and calibration mostly employ traditional morphological processing (such as dilation and erosion) or single threshold segmentation methods. These methods can only perform simple screening at the pixel grayscale level and cannot capture the spatial correlation features of defect regions, making it difficult to distinguish between small defects with blurred boundaries and normal circuits. In contrast, the defect region calibration network of this invention innovatively introduces graph convolutional networks to mine spatial correlation features, enhances effective features through the ReLU function, and achieves accurate calibration using defect binary masks. Furthermore, it reuses the defect enhancement features generated in the early stage and the dynamic segmentation threshold after illumination calibration throughout the process. This not only overcomes the limitations of existing technologies that only focus on grayscale features and ignore spatial correlation, but also solves core problems such as too many false defects, blurred defect boundaries, and lack of targeted calibration, enabling the output of more accurate original defect candidate region features.

[0014] Furthermore, step A3 also includes: The coupled multidimensional defect features are subjected to feature neighborhood offset processing and convolution operation. The local feature consistency score is obtained through feature fusion and global pooling operation. Then, the coupled multidimensional defect features are fused together. The effective defect features are retained by residual connection and the interference of normal line texture is removed to obtain the interference-free pure defect features.

[0015] Furthermore, step A3 also includes: A31: Based on the optimized defect candidate features and the cleaned circuit board optical image, multidimensional initial defect features are extracted. The calculation method is as follows: in, For texture features, For a 3×3 depthwise separable convolution, These are the optimized candidate features for defects. For multidimensional initial defect characteristics, For attention mechanisms, To add element by element, It is a 5×5 convolutional layer. It is a multilayer perceptron. Optical image of the circuit board after cleaning; A32: Based on the initial features of the multidimensional defect and the illumination-aware attention weights, multi-scale features are fused and key information is enhanced to obtain the coupled multidimensional defect features. The calculation method is as follows: in, For multi-scale fusion features, For upsampling, It is a 3×3 convolutional layer. For Hadama accumulation, For the multidimensional defect characteristics after coupling, For global average pooling, Light perception attention weights; A33: Based on the coupled multidimensional defect characteristics, a texture suppression network is used to remove interference from normal line textures, resulting in clean defect characteristics after interference removal. The texture suppression network is calculated as follows: in, For local feature consistency score, To offset the feature by 1 pixel in both the horizontal and vertical directions, These are the pure defect features after removing interference. For the Sigmoid function, For residual connection layer; A34: Based on the clean defect characteristics after interference removal, extract the subtle short-circuit defect characteristics of the circuit board. The calculation method is as follows: in, Preliminary characteristics of minor short-circuit defects on circuit boards. It is a fully connected layer. For layer normalization, This is a characteristic of minor short-circuit defects on circuit boards. Calculated for the mean. This refers to the Softmax function.

[0016] It should be further explained that the coupled multidimensional defect features are still mixed with interference from normal line textures. The local feature distribution of these normal line textures is highly similar to that of the minor short-circuit defect features, and the two are intertwined in spatial structure, making them difficult to distinguish accurately. At the same time, a small number of pseudo-features caused by illumination fluctuations remain in the coupled features, causing the defect features to be masked by interference information. In addition, the minor short-circuit defects themselves are very small in size, and their features are easily removed along with the interference information during the process of suppressing interference, resulting in the loss of defect features. The texture suppression network constructs a complete logical closed loop from interference feature identification and accurate suppression to defect feature preservation through the coordinated linkage of feature neighborhood offset, convolution operation, local feature consistency calculation, interference suppression weight generation and residual connection. Each link is progressive and mutually supportive. First, to address the difficulty in identifying the differences between normal line textures and defect features in local areas, a feature neighborhood offset operation is used to offset the coupled multidimensional defect features by 1 pixel in both the horizontal and vertical directions, obtaining neighborhood comparison data for the defect features. Then, convolution processing is performed on the coupled multidimensional defect features and the offset features, and Hadamard product operation is used to enhance the feature differences in local areas. Finally, global average pooling is used to integrate them to obtain a local feature consistency score. Normal line textures, due to their regular structure, have a significantly higher local feature consistency score than irregular, subtle short-circuit defects, thus achieving a preliminary distinction between interference features and defect features. Second, the difference between the local feature consistency score and 1 is calculated, and this difference is input into the multilayer sensing... The system performs feature transformation and then activates the Sigmoid function to generate interference suppression weights. These weights correspond to normal circuit texture areas, providing high-intensity suppression for them while maintaining low-intensity suppression for defect areas, thus achieving accurate screening of interference features. Finally, the interference suppression weights are fused with the coupled multidimensional defect features to remove interference from normal circuit textures. At the same time, the residual connection layer preserves the effective defect details in the coupled multidimensional defect features, avoiding the loss of defect features during the interference suppression process. This ensures that the output clean defect features after interference removal are free of redundant interference and fully retain the core features of subtle short-circuit defects, providing a high-quality feature foundation for the subsequent extraction of subtle short-circuit defect features of circuit boards. Existing technologies for handling texture interference in circuit board defect features mostly employ global threshold removal, single-dimensional convolutional denoising, or simple feature subtraction. These methods either only coarsely process obvious interference globally, failing to accurately identify and suppress normal circuit textures similar to defect features, resulting in incomplete interference removal; or they remove some minor defect features along with the interference suppression, leading to incomplete defect features. Furthermore, existing technologies lack targeted processing mechanisms for local feature differences, making it difficult to cope with complex and varied texture interference scenarios on circuit boards, resulting in poor processing stability. In contrast, the texture suppression network of this invention achieves accurate differentiation between interference features and defect features based on local feature consistency scores, achieves targeted removal of interference through interference suppression weights, and achieves complete preservation of defect features through residual connections. This not only overcomes the shortcomings of existing technologies such as inaccurate interference removal and easy loss of defect features, but also adapts to the texture distribution characteristics of different circuit boards, resulting in more stable and accurate processing effects and significantly improving the purity and completeness of defect features.

[0017] Furthermore, step A4 also includes: A41: Based on the optimized defect candidate features and the circuit board's minor short-circuit defect features, feature fusion and weight optimization are completed to obtain the weight-optimized fused features. The calculation method is as follows: in, A 1×1 convolutional layer, For candidate region location features, These are the fusion features after weight optimization; A42: Based on the fusion features after weight optimization, calculate the calibrated defect confidence level. The calculation method is as follows: in, The confidence level of the original defect. The calibrated defect confidence level. This is the illumination calibration factor. To calibrate the base offset; A43: Based on the calibrated defect confidence level, the effective defect region mask is obtained through threshold filtering. The calculation method is as follows: in, This is a mask for the effective defect area.

[0018] Furthermore, step A5 also includes: A51: The effective defect area mask is used to filter the effective areas of the minor short-circuit defect features of the circuit board. The areas with a mask value of 1 are retained as effective defect candidate areas, and the invalid areas with a mask value of 0 are removed. A52: Based on the pixel coordinates of the effective defect candidate region in the feature of minor short-circuit defects on the circuit board, perform inverse mapping of pixel coordinates to restore the original pixel coordinates of the effective defect candidate region in the optical image of the cleaned circuit board, and determine the vertex coordinates of the bounding box of the defect region. A53: Calculate the pixel length and width of the defect area based on the vertex coordinates of the bounding box of the defect area, and combine it with the scale of the optical image of the cleaned circuit board to obtain the actual physical size of the defect. A54: Integrate the boundary box vertex coordinates of the defect region and the actual physical size of the defect to obtain the detection results of minor short-circuit defects on the circuit board.

[0019] This invention also discloses a machine vision-based circuit board micro-short circuit detection system, comprising: Circuit board image acquisition module: Acquires raw optical image data and light intensity data of circuit board image, and cleans them to obtain cleaned optical image and light intensity data of circuit board image; Defect candidate feature extraction module: Based on the optical image of the cleaned circuit board, the defect enhancement features are calculated through the micro-defect enhancement network; combined with the light intensity data of the cleaned circuit board, the light perception attention weight and the dynamic segmentation threshold after light calibration are calculated; the defect region is segmented and calibrated through the defect region calibration network to obtain the original defect candidate region features, and then the optimized defect candidate features are extracted. The circuit board minute short circuit defect feature extraction module extracts multi-dimensional initial defect features based on optimized defect candidate features and cleaned circuit board optical images; it then extracts coupled multi-dimensional defect features by combining illumination perception attention weights; and finally extracts clean defect features after interference removal through a texture suppression network, thereby extracting the circuit board minute short circuit defect features. Effective defect region calculation module: Based on the optimized defect candidate features and the circuit board minor short circuit defect features, extract the weighted fusion features, calculate the calibrated defect confidence, and obtain the effective defect region mask through threshold filtering; Defect detection module: It filters the effective defect area through the effective defect area mask, determines the coordinates of the boundary box vertices of the defect area, and then calculates the actual physical size of the defect to obtain the detection results of minor short circuit defects on the circuit board.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention first collects and cleans relevant images and lighting data of circuit boards. After a series of feature enhancement, lighting adaptation, interference removal and feature fusion processing, the final detection result is obtained through confidence calculation and threshold screening. It effectively solves the problems of low efficiency and high false negative rate of traditional manual detection, poor resistance to lighting interference of traditional machine vision, and insufficient targeting of existing deep learning solutions for minor short circuit defects. It meets the actual production needs of the electronics manufacturing industry for high-quality and high-efficiency detection of circuit boards.

[0021] (2) In view of the problem that the micro short-circuit defects on the circuit board are small in size and have very little difference in gray level from the background information such as substrate texture, and that the effective defect details are easily lost in the feature processing of the cleaned image, resulting in the defect features being submerged and hindering subsequent segmentation and calibration, this invention constructs a micro-defect enhancement network. Through multi-scale convolution processing, feature activation operation and residual connection work together, multi-scale convolution fully captures local defects, mid-layer texture and global background features, feature activation operation dynamically amplifies the difference between defects and background, and residual connection retains the original defect details to the maximum extent and avoids the omission of key information.

[0022] (3) In view of the fact that the optical image of the cleaned circuit board is easily affected by the light fluctuation, the difference between the gray scale of the minor short-circuit defects and the substrate texture is very small, and the light change will amplify the feature confusion and cause gray scale distortion, this invention designs light-sensing attention weights and dynamic segmentation thresholds after light calibration. First, the dimension of the light intensity data is adjusted by dimensional expansion so that it is spliced ​​with the defect enhancement features, taking into account both defect features and light factors. Then, the fused data is convolved and the light-sensing attention weights are obtained by the Sigmoid function to accurately mark the light abnormal area. Finally, the defect enhancement features are processed by global average pooling and fused with the attention weights to generate a dynamic segmentation threshold. Different adjustments are made for different light areas to achieve light adaptation and accurate threshold generation. This solves the problem of fusion of light data and image features, realizes dynamic adaptation of segmentation thresholds, and improves the accuracy of defect area segmentation.

[0023] (4) To address the problem of defect features being masked by interference from normal line textures and residual illumination pseudo-features, this invention constructs a texture suppression network. First, it strengthens the differences in local features through feature neighborhood offset and convolution operations. Then, it obtains a local feature consistency score through global average pooling to initially distinguish between interference and defect features. Next, it generates interference suppression weights by passing the score difference through a multilayer perceptron and a sigmoid function to accurately suppress normal line textures. Finally, it fuses the weights to remove interference and retains effective defect details through residual connections, outputting pure defect features and providing a high-quality foundation for subsequent defect feature extraction. Attached Figure Description

[0024] Figure 1 A flowchart illustrating a machine vision-based method for detecting minute short circuits on circuit boards provided by this invention. Figure 2 The feature map for the defect enhancement feature provided by the present invention. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0026] Example 1: A machine vision-based method for detecting minute short circuits on circuit boards, such as... Figure 1 As shown, it includes the following steps: A1: Acquire the original optical image data and illumination intensity data of the circuit board, and perform cleaning to obtain the cleaned optical image and illumination intensity data of the circuit board, including: The original optical image data of the circuit board is acquired by using a pixel industrial CCD camera with a ring-shaped LED uniform light source. The data type is two-dimensional grayscale pixel data. The optical image includes the front surface image and the back surface image of the circuit board. The original optical image data of the circuit board is preprocessed by Gaussian filtering noise reduction method to obtain the cleaned optical image of the circuit board. The illumination intensity data of the circuit board image is collected by an ambient light sensor. The data type is numerical data, which includes the ambient light intensity value at the time of collection. Then, the illumination intensity data of the circuit board after cleaning is obtained by removing constant values ​​of three times standard difference and smoothing by sliding window mean.

[0027] A2: Based on the optical image of the cleaned circuit board, a micro-defect enhancement network is used to calculate defect enhancement features, thereby enhancing the contrast between the region of subtle short-circuit defects and the background to highlight the defect features and solving the problem that subtle short-circuit defect features are easily masked by background information in existing solutions. Combined with the illumination intensity data of the cleaned circuit board, illumination perception attention weights and dynamic segmentation thresholds after illumination calibration are calculated. A defect region calibration network is used for defect region segmentation and calibration to obtain the original defect candidate region features, and then optimized defect candidate features are extracted, including: A21: Based on the optical image of the cleaned circuit board, the defect enhancement features are calculated using a micro-defect enhancement network. The calculation method of the micro-defect enhancement network is as follows: in, For multi-scale convolutional fusion features, A 1×1 convolutional layer, This is an optical image of the circuit board after cleaning. To add element by element, It is a 3×3 convolutional layer. It is a 5×5 convolutional layer. For defect enhancement features, such as Figure 2 As shown, For Hadama accumulation, For the Sigmoid function, It is a multilayer perceptron. For residual connection layer; A22: Based on the defect enhancement features and the illumination intensity data of the cleaned circuit board, calculate the illumination perception attention weight and the dynamic segmentation threshold after illumination calibration. The calculation method is as follows: in, For light perception attention weights, For splicing operations, For dimensional expansion operations, The data shows the light intensity of the circuit board after cleaning. This is the dynamic segmentation threshold after illumination calibration. For global average pooling; A23: Based on the defect enhancement features, illumination-aware attention weights, and the dynamic segmentation threshold after illumination calibration, defect region segmentation and calibration are performed using a defect region calibration network to obtain the original defect candidate region features. The calculation method of the defect region calibration network is as follows: in, For defect binary mask, For ReLU function, For graph convolutional networks, For Boolean matrix generation operations, Candidate features for original defects; A24: Based on the original candidate defect features, extract the optimized candidate defect features. The calculation method is as follows: in, Weights are used to verify the validity of candidate regions. For gating attention mechanism, These are the optimized candidate features for defects.

[0028] A3: Based on the optimized defect candidate features and the cleaned optical image of the circuit board, multidimensional initial defect features are extracted; combined with illumination-aware attention weights, coupled multidimensional defect features are extracted; then, through a texture suppression network, the purified defect features after interference removal are extracted, and finally, the subtle short-circuit defect features of the circuit board are extracted, including: A31: Based on the optimized defect candidate features and the cleaned circuit board optical image, multidimensional initial defect features are extracted. The calculation method is as follows: in, For texture features, For a 3×3 depthwise separable convolution, These are the optimized candidate features for defects. For multidimensional initial defect characteristics, For attention mechanisms, To add element by element, It is a 5×5 convolutional layer. It is a multilayer perceptron. Optical image of the circuit board after cleaning; A32: Based on the initial features of the multidimensional defect and the illumination-aware attention weights, multi-scale features are fused and key information is enhanced to obtain the coupled multidimensional defect features. The calculation method is as follows: in, For multi-scale fusion features, For upsampling, It is a 3×3 convolutional layer. For Hadama accumulation, For the multidimensional defect characteristics after coupling, For global average pooling, Light perception attention weights; A33: Based on the coupled multidimensional defect characteristics, a texture suppression network is used to remove interference from normal line textures, resulting in clean defect characteristics after interference removal. The texture suppression network is calculated as follows: in, For local feature consistency score, To offset the feature by 1 pixel in both the horizontal and vertical directions, These are the pure defect features after removing interference. For the Sigmoid function, For residual connection layer; A34: Based on the clean defect characteristics after interference removal, extract the subtle short-circuit defect characteristics of the circuit board. The calculation method is as follows: in, Preliminary characteristics of minor short-circuit defects on circuit boards. It is a fully connected layer. For layer normalization, This is a characteristic of minor short-circuit defects on circuit boards. Calculated for the mean. This refers to the Softmax function.

[0029] Specifically, for scenarios where the distribution of subtle short-circuit defect features is uneven and there is local feature redundancy in high-density circuit boards, this invention also provides a calculation method to enhance local key defect features and suppress redundant information, replacing step A34. The calculation method is as follows: in, For local response normalization operation, This represents the local response normalization feature.

[0030] A4: Based on the optimized candidate defect features and the circuit board's minor short-circuit defect features, extract the weighted fusion features, calculate the calibrated defect confidence, and obtain the effective defect region mask through threshold filtering, including: A41: Based on the optimized defect candidate features and the circuit board's minor short-circuit defect features, feature fusion and weight optimization are completed to obtain the weight-optimized fused features. The calculation method is as follows: in, A 1×1 convolutional layer, For candidate region location features, These are the fusion features after weight optimization; A42: Based on the fusion features after weight optimization, calculate the calibrated defect confidence level. The calculation method is as follows: in, The confidence level of the original defect. The calibrated defect confidence level. This is the illumination calibration factor. To calibrate the base offset; A43: Based on the calibrated defect confidence level, the effective defect region mask is obtained through threshold filtering. The calculation method is as follows: in, This is a mask for the effective defect area.

[0031] A5: Effective defect areas are filtered using a mask, and the coordinates of the boundary box vertices of the defect areas are determined. The actual physical dimensions of the defects are then calculated to obtain the detection results for minor short-circuit defects on the circuit board, including: A51: The effective defect area mask is used to filter the effective areas of the minor short-circuit defect features of the circuit board. The areas with a mask value of 1 are retained as effective defect candidate areas, and the invalid areas with a mask value of 0 are removed. A52: Based on the pixel coordinates of the effective defect candidate region in the feature of minor short-circuit defects on the circuit board, perform inverse mapping of pixel coordinates to restore the original pixel coordinates of the effective defect candidate region in the optical image of the cleaned circuit board, and determine the vertex coordinates of the bounding box of the defect region. A53: Calculate the pixel length and width of the defect area based on the vertex coordinates of the bounding box of the defect area, and combine it with the scale of the optical image of the cleaned circuit board to obtain the actual physical size of the defect. A54: Integrate the boundary box vertex coordinates of the defect region and the actual physical size of the defect to obtain the detection results of minor short-circuit defects on the circuit board.

[0032] This embodiment is applied to the mass production inspection scenario of high-density FR-4 PCB circuit boards in the consumer electronics field. The light intensity fluctuation range in the production workshop is 500-1200 lux. The image scale is set to 1 pixel = 0.01 mm, the light calibration coefficient α = 0.3, the calibration base offset β = 0.05, and the defect confidence threshold is 0.5. After acquiring and cleaning the circuit board image, it undergoes preliminary feature enhancement, illumination adaptation, and texture suppression. Step A4 then completes feature fusion, weight optimization, and confidence calibration to generate an effective defect area mask. Based on this mask, step A5 sequentially performs effective area filtering, pixel coordinate inverse mapping, and physical size conversion. Finally, step A54 integrates the results to obtain the complete detection outcome, as detailed below: Defect bounding box vertex coordinates (corresponding to the pixel coordinates of the optical image of the cleaned circuit board): top left corner (320, 250), top right corner (380, 250), bottom right corner (380, 310), bottom left corner (320, 310); Actual physical dimensions of the defect: Based on the image scale of 1 pixel = 0.01 mm, the length of the defect is (380-320) × 0.01 mm = 0.6 mm, and the width is (310-250) × 0.01 mm = 0.6 mm. The actual dimensions of the defect are 0.6 mm × 0.6 mm. Defect confidence level: The calibrated defect confidence level is 0.5308, which is higher than the set threshold of 0.5, and is therefore determined to be a valid defect; Traditional fixed threshold segmentation method: Without introducing an illumination calibration mechanism, the original defect confidence score is directly calculated to be 0.48. Since it is below the threshold of 0.5, it is judged as no defect, resulting in missed detection. The core reason is that 620 lux illumination distorts the grayscale features of the defect. The fixed threshold cannot be adapted to dynamic lighting scenarios and cannot correct the confidence score deviation caused by illumination interference.

[0033] In this invention, all neural network modules, including the micro-defect enhancement network, defect region calibration network, and texture suppression network, are trained end-to-end using a unified training framework. The training dataset is constructed based on the image acquisition method in the A1 step. The dataset contains manually labeled optical images of circuit boards under different lighting conditions, with different defect types and sizes, along with corresponding lighting intensity data. The labeled dataset is divided into training, validation, and test sets according to a preset ratio for network training, parameter adjustment, and performance verification. During training, the cross-entropy loss function is used as the core loss function, combined with the mean squared error loss function to constrain the feature bias in the defect feature extraction process. The final total loss function is obtained through weighted summation, achieving dual optimization of defect recognition accuracy and feature extraction accuracy. The optimizer used is the Adam optimizer, which adaptively adjusts the learning rate. To balance training speed and convergence stability, the initial learning rate is set using a commonly used default baseline value. During training, the learning rate is dynamically adjusted using a learning rate decay strategy based on the loss changes on the validation set, avoiding gradient vanishing or overfitting. To further improve the model's generalization ability, conventional regularization methods such as weight decay and Dropout are used during training to suppress overfitting of network parameters to the training set data. A fixed number of training iterations and an early stopping mechanism are also set; training is automatically terminated when the loss value on the validation set fails to decrease for several consecutive iterations, ensuring that the trained network model has both good detection accuracy and operational stability. After training, the overall performance of the network model is validated using a test set. Once the model meets the preset requirements in terms of defect identification accuracy, false negative rate, and false positive rate, it can be deployed for practical applications in detecting minor short-circuit defects on circuit boards.

[0034] Example 2: This invention also discloses a machine vision-based circuit board micro-short circuit detection system, comprising: Circuit board image acquisition module: Acquires raw optical image data and light intensity data of circuit board image, and cleans them to obtain cleaned optical image and light intensity data of circuit board image; Defect candidate feature extraction module: Based on the optical image of the cleaned circuit board, the defect enhancement features are calculated through the micro-defect enhancement network; combined with the light intensity data of the cleaned circuit board, the light perception attention weight and the dynamic segmentation threshold after light calibration are calculated; the defect region is segmented and calibrated through the defect region calibration network to obtain the original defect candidate region features, and then the optimized defect candidate features are extracted. The circuit board minute short circuit defect feature extraction module extracts multi-dimensional initial defect features based on optimized defect candidate features and cleaned circuit board optical images; it then extracts coupled multi-dimensional defect features by combining illumination perception attention weights; and finally extracts clean defect features after interference removal through a texture suppression network, thereby extracting the circuit board minute short circuit defect features. Effective defect region calculation module: Based on the optimized defect candidate features and the circuit board minor short circuit defect features, extract the weighted fusion features, calculate the calibrated defect confidence, and obtain the effective defect region mask through threshold filtering; Defect detection module: It filters the effective defect area through the effective defect area mask, determines the coordinates of the boundary box vertices of the defect area, and then calculates the actual physical size of the defect to obtain the detection results of minor short circuit defects on the circuit board.

[0035] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0036] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0037] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for detecting a fine short circuit of a circuit board based on machine vision, characterized by, Includes the following steps: A1: Collect the original optical image data and illumination intensity data of the circuit board, and clean them to obtain the cleaned optical image and illumination intensity data of the circuit board. A2: Based on the optical image of the cleaned circuit board, the defect enhancement features are calculated using a micro-defect enhancement network; combined with the light intensity data of the cleaned circuit board, the light perception attention weight and the dynamic segmentation threshold after light calibration are calculated. Defect region segmentation and calibration are performed using a defect region calibration network to obtain the original defect candidate region features, and then the optimized defect candidate features are extracted. Step A2 includes: Based on defect enhancement features and illumination intensity data of the cleaned circuit board, the illumination perception attention weight and the dynamic segmentation threshold after illumination calibration are calculated as follows: wherein, is an illumination-aware attention weight, is is a sigmoid function, is a concatenation operation, is a dimension expansion operation, is the cleaned-up board illumination intensity data, is the illumination-calibrated dynamic segmentation threshold, is a 3x3 size convolution layer, is a global average pooling, is a defect-enhanced feature, is a Hadamard product; Based on defect enhancement features, illumination-aware attention weights, and dynamic segmentation thresholds after illumination calibration, defect region segmentation and calibration are performed using a defect region calibration network to obtain original defect candidate region features. The calculation method of the defect region calibration network is as follows: wherein, is a defect binary mask, is a ReLU function, is a graph convolution network, is a Boolean matrix generation operation, is a raw defect candidate feature; A3: extracting a multi-dimensional initial defect feature according to the optimized defect candidate feature and the cleaned circuit board optical image; extracting a coupled multi-dimensional defect feature in combination with the illumination perception attention weight; then extracting a pure defect feature after interference suppression through a texture suppression network, and further extracting a circuit board fine short circuit defect feature; Step A3 includes: Based on the coupled multidimensional defect features, a texture suppression network is used to remove interference from normal circuit textures, resulting in clean defect features after interference removal. The texture suppression network is calculated as follows: in, For local feature consistency score, For the multidimensional defect characteristics after coupling, To offset the feature by 1 pixel in both the horizontal and vertical directions, These are the pure defect features after removing interference. For the Sigmoid function, , A4: Based on the optimized defect candidate features and the circuit board minor short-circuit defect features, extract the weighted fusion features, calculate the calibrated defect confidence, and obtain the effective defect area mask through threshold screening. A5: Valid regions are filtered by using a mask of valid defect regions, and the coordinates of the vertices of the bounding box of the defect region are determined. Then, the actual physical size of the defect is calculated to obtain the detection results of minor short-circuit defects on the circuit board.

2. The method for detecting minute short circuits on circuit boards based on machine vision according to claim 1, characterized in that, Step A1 includes: The original optical image data of the circuit board is acquired by using a pixel industrial CCD camera with a ring-shaped LED uniform light source. The data type is two-dimensional grayscale pixel data. The optical image includes the front surface image and the back surface image of the circuit board. The original optical image data of the circuit board is preprocessed by Gaussian filtering noise reduction method to obtain the cleaned optical image of the circuit board. The illumination intensity data of the circuit board image is collected by an ambient light sensor. The data type is numerical data, which includes the ambient light intensity value at the time of collection. Then, the illumination intensity data of the circuit board after cleaning is obtained by removing constant values ​​of three times standard difference and smoothing by sliding window mean.

3. The method for detecting minute short circuits on circuit boards based on machine vision according to claim 1, characterized in that, The micro-defect enhancement network in step A2 includes: The cleaned circuit board optical image is processed by multi-scale convolution through convolutional layers of size 1×1, 3×3, and 5×5, respectively. The results are then added element-wise to obtain the multi-scale convolutional fusion feature. After processing the cleaned circuit board optical image through a multilayer perceptron, the result calculated by the Sigmoid function is combined with the multi-scale convolutional fusion feature by a Hadamard product operation. Finally, this result is added to the result of the cleaned circuit board optical image after processing by the residual connection layer to obtain the defect enhancement feature.

4. The method for detecting minute short circuits on circuit boards based on machine vision according to claim 3, characterized in that, The defect region calibration network in step A2 includes: The defect enhancement features are calculated using a graph convolutional network and the ReLU function, and then a Boolean matrix generation operation is performed with the dynamic segmentation threshold after illumination calibration. The defect binary mask is then obtained through Hadamard product operation. The defect enhancement features and the defect binary mask are then subjected to Hadamard product operation to complete the defect region segmentation and calibration, and the original defect candidate region features are obtained.

5. The method for detecting minute short circuits on circuit boards based on machine vision according to claim 3, characterized in that, Step A2 includes: Based on the optical image of the cleaned circuit board, defect enhancement features are calculated using a micro-defect enhancement network. The calculation method of the micro-defect enhancement network is as follows: in, For multi-scale convolutional fusion features, A 1×1 convolutional layer, This is an optical image of the circuit board after cleaning. To add element by element, For residual connection layer; Based on defect enhancement features and illumination intensity data of the cleaned circuit board, the illumination perception attention weight and the dynamic segmentation threshold after illumination calibration are calculated as follows: in, For light perception attention weights, For splicing operations, For dimensional expansion operations, The data shows the light intensity of the circuit board after cleaning. This is the dynamic segmentation threshold after illumination calibration. For global average pooling; Based on the original candidate defect features, optimized candidate defect features are extracted, calculated as follows: in, Weights are used to verify the validity of candidate regions. For gating attention mechanism, These are the optimized candidate features for defects.

6. The method for detecting minute short circuits on circuit boards based on machine vision according to claim 3, characterized in that, Step A3 includes: The coupled multidimensional defect features are subjected to feature neighborhood offset processing and convolution operation. The local feature consistency score is obtained through feature fusion and global pooling operation. Then, the coupled multidimensional defect features are fused together. The effective defect features are retained by residual connection and the interference of normal line texture is removed to obtain the interference-free pure defect features.

7. The method for detecting minute short circuits on circuit boards based on machine vision according to claim 6, characterized in that, Step A3 includes: Based on the optimized defect candidate features and the cleaned optical image of the circuit board, multidimensional initial defect features are extracted. The calculation method is as follows: in, For texture features, For a 3×3 depthwise separable convolution, These are the optimized candidate features for defects. For multidimensional initial defect characteristics, For attention mechanisms, To add element by element, It is a 5×5 convolutional layer. It is a multilayer perceptron. Optical image of the circuit board after cleaning; Based on the initial features of the multidimensional defects and the illumination-sensing attention weights, multi-scale features are fused and key information is enhanced to obtain the coupled multidimensional defect features. The calculation method is as follows: in, For multi-scale fusion features, For upsampling, It is a 3×3 convolutional layer. For Hadama accumulation, For global average pooling, Light perception attention weights; Based on the clean defect characteristics after interference removal, the subtle short-circuit defect characteristics of the circuit board are extracted. The calculation method is as follows: in, Preliminary characteristics of minor short-circuit defects on circuit boards. It is a fully connected layer. For layer normalization, This is a characteristic of minor short-circuit defects on circuit boards. Calculated for the mean. for oftmax function.

8. The method for detecting minute short circuits on circuit boards based on machine vision according to claim 7, characterized in that, The A4 step includes: A41: Based on the optimized defect candidate features and the circuit board's minor short-circuit defect features, feature fusion and weight optimization are completed to obtain the weight-optimized fused features. The calculation method is as follows: in, A 1×1 convolutional layer, For candidate region location features, These are the fusion features after weight optimization; A42: Based on the fusion features after weight optimization, calculate the calibrated defect confidence level. The calculation method is as follows: in, The confidence level of the original defect. The calibrated defect confidence level. This is the illumination calibration factor. To calibrate the base offset; A43: Based on the calibrated defect confidence level, the effective defect region mask is obtained through threshold filtering. The calculation method is as follows: in, This is a mask for the effective defect area.

9. The method for detecting minute short circuits on circuit boards based on machine vision according to claim 8, characterized in that, Step A5 includes: A51: The effective defect area mask is used to filter the effective areas of the minor short-circuit defect features of the circuit board. The areas with a mask value of 1 are retained as effective defect candidate areas, and the invalid areas with a mask value of 0 are removed. A52: Based on the pixel coordinates of the effective defect candidate region in the feature of minor short-circuit defects on the circuit board, perform inverse mapping of pixel coordinates to restore the original pixel coordinates of the effective defect candidate region in the optical image of the cleaned circuit board, and determine the vertex coordinates of the bounding box of the defect region. A53: Calculate the pixel length and width of the defect area based on the vertex coordinates of the bounding box of the defect area, and combine it with the scale of the optical image of the cleaned circuit board to obtain the actual physical size of the defect. A54: Integrate the boundary box vertex coordinates of the defect region and the actual physical size of the defect to obtain the detection results of minor short-circuit defects on the circuit board.

10. A machine vision-based system for detecting minute short circuits on circuit boards, characterized in that, include: Circuit board image acquisition module: Acquires raw optical image data and light intensity data of circuit board image, and cleans them to obtain cleaned optical image and light intensity data of circuit board image; Defect candidate feature extraction module: Based on the optical image of the cleaned circuit board, the defect enhancement features are calculated through the micro-defect enhancement network; combined with the light intensity data of the cleaned circuit board, the light perception attention weight and the dynamic segmentation threshold after light calibration are calculated. Defect region segmentation and calibration are performed using a defect region calibration network to obtain the original defect candidate region features, and then the optimized defect candidate features are extracted. Circuit board minor short-circuit defect feature extraction module: Extracts multi-dimensional initial defect features based on optimized defect candidate features and cleaned circuit board optical images; By combining illumination perception attention weights, multidimensional defect features are extracted after coupling. Then, a texture suppression network is used to extract the clean defect features after interference removal, and then the subtle short-circuit defect features of the circuit board are extracted. Effective defect region calculation module: Based on the optimized defect candidate features and the circuit board minor short circuit defect features, extract the weighted fusion features, calculate the calibrated defect confidence, and obtain the effective defect region mask through threshold filtering; Defect detection module: Valid regions are filtered by using a mask of valid defect regions, and the coordinates of the vertices of the bounding box of the defect region are determined. The actual physical size of the defect is then calculated to obtain the detection result of minor short circuit defects on the circuit board; thereby realizing the machine vision-based minor short circuit detection method for circuit boards as described in any one of claims 1-9.

Images