A PCB defect detection system and method based on an improved YOLO algorithm

Abstract

This invention relates to a PCB defect detection system and method based on an improved YOLO algorithm. The detection system employs an improved YOLO algorithm, optimizing data augmentation to reduce the instability caused by single-mosaic data augmentation and improving model accuracy. The system includes: an image acquisition and processing module, a neural network model improvement module, a model training module, a model validation module, and a defect detection module. The detection method involves the following steps: building a PCB defect dataset image acquisition system; acquiring PCB surface defect images using an industrial CCD camera as the raw dataset; preprocessing the dataset images; improving the YOLO network model; training the model parameters; generating a PCB defect detection model; and testing the test set. The defect detection system built by this invention can achieve automatic PCB defect detection, offering advantages such as low cost, high speed, and high accuracy.

Description

Technical Field

[0001] This invention relates to the field of computer vision defect detection technology, and in particular to a PCB defect detection system and method based on an improved YOLO algorithm. Background Technology

[0002] Printed Circuit Boards (PCBs) are a crucial component of modern electronic devices, used to connect and support electronic components. With the development of electronic manufacturing, the trend towards miniaturization and increasing complexity of PCBs is inevitable. This leads to a greater likelihood of various defects during the PCB manufacturing process, such as vias, rodent bites, open circuits, short circuits, burrs, and fake copper. These defects can cause circuit malfunctions or performance degradation, and in severe cases, may even endanger the lives of users. However, most PCB defects are very small and their color is similar to the background color of the PCB traces, making them very difficult to detect with the naked eye, and the detection accuracy is also low. Therefore, developing efficient PCB defect detection technology is of great significance in modern electronics manufacturing, as it can effectively improve product quality, reduce production costs, and ensure the stability and reliability of electronic devices.

[0003] Currently, mainstream PCB defect detection methods include machine vision inspection, manual sampling, X-ray inspection, and ultrasonic testing. Manual sampling is inefficient and prone to missed or false positives. X-ray inspection offers high accuracy but is costly and slow, making it suitable only for products with very high precision requirements. Ultrasonic testing can only detect internal PCB defects and cannot detect surface problems. For PCB surface defect detection, machine vision inspection is cost-effective and efficient, making it the most mainstream method currently.

[0004] Object detection algorithms are currently mainly divided into two categories: single-object detection algorithms, spearheaded by YOLO, and dual-object detection algorithms, spearheaded by Faster R-CNN. Among these, dual-object detection algorithms suffer from poor model accuracy and slow recognition speed when dealing with imbalanced defect types or small defect sizes. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a PCB defect detection system based on an improved YOLO algorithm. The detection system employs an improved YOLOX algorithm, which optimizes data augmentation, reduces the instability caused by single Mosaic data augmentation, and improves the accuracy of the model. The system includes: an image acquisition and processing module, a neural network model improvement module, a model training module, a model verification module, and a defect detection module.

[0006] The image acquisition and processing module acquires images of PCB surface defects by building an image acquisition system and processes the acquired images to obtain a dataset. The image acquisition and processing module includes a camera, a lens, a clamping device, and a base. A ring light source is also provided on the lens side. After the camera and lens are assembled, they are placed together on the clamping device. The camera's output interface is connected to the computer hardware.

[0007] The neural network model improvement module, model training module, model validation module, and defect detection module all run on computer hardware as data modules. The neural network model improvement module is based on the YOLOX neural network and is used to build a defect detection model. The improvements include using a focus structure to stack feature layers, expanding the input channels; reconstructing CSPLAYer with ResBlock in the feature fusion stage, the reconstructed CSPLAY module can better handle multi-scale feature fusion, obtain more semantic information, and alleviate the gradient repetition problem; then, spatial pyramid pooling is used to process max pooling at four different scales, which greatly improves the receptive field; finally, a decoupled head is used as the YOLO head.

[0008] The model training module is used to set the network parameters of the defect detection model. The created dataset is input into the defect detection model for training. Adam is used as the optimizer during training. The learning rate of Adam optimizer is set to 0.001, the weight decay is set to 0.0005, the number of training epochs is set to 500, and the batch size is set to 4.

[0009] The model validation module is used to test the initially trained defect detection model with a validation set, and adjust the learning rate and epoch parameters according to the validation results to improve the validation accuracy and obtain the final defect detection model.

[0010] The defect detection module is used to input test set data into the trained defect detection model, detect and label the type, location and confidence level of PCB surface defects in the test set.

[0011] In one embodiment of the present invention, the light-collecting lens ring on the lens and the PCB sample to be tested on the base are on the same vertical straight line, and the base is also provided with a rectangular light source and a light source controller. At the same time, the input terminal of the light source controller is connected to the computer hardware by a control line. The camera is an industrial CCD camera, and a ring light source is coaxially and vertically fixed above the PCB board to be tested. A rectangular light source is placed below the PCB board. The defect image of the PCB surface is obtained by adjusting the distance between the camera, the light source and the PCB board and the exposure time.

[0012] This invention also provides a PCB defect detection method based on an improved YOLO algorithm, which, based on the aforementioned PCB defect detection system, includes the following steps:

[0013] Step S1: Set up an image acquisition system to acquire images of defective samples;

[0014] Step S2: Edit the acquired defect images to obtain images with uniform dimensions;

[0015] Step S3: Use data augmentation methods to augment the samples, and then perform noise reduction and contrast enhancement on the images;

[0016] Step S4: Select image samples from the images, label the location and type of defects in the image samples, and create a dataset;

[0017] Step S5: Input the processed RGB three-channel image into the improved YOLOX neural network for training to obtain a target detection model for PCB panel defects;

[0018] Step S6: Read the obtained target detection model and test it on the test set prepared in step S4 to realize the identification, classification and location of PCB defects;

[0019] The target detection model in specific step S5 includes: a CSPDarknet53 backbone feature extraction network with a focus structure, a spatial pyramid pooling structure SPP, a path aggregation network PAPPN, and a YOLO Head with a decoupled head.

[0020] In one embodiment of the present invention, the size of the image after unification in step S2 is 640*640.

[0021] In one embodiment of the present invention, the data augmentation method described in step S3 specifically involves first performing translation, flipping, and rotation on the acquired defect image to augment the data, and then performing image denoising and enhancement on the augmented image. First, a combination of average filtering and smooth filtering is used to perform denoising filtering on the image, wherein moving is selected as the smoothing method for smoothing filtering, and the span is set to 5; the window number for average filtering is set to 4. This denoising method can effectively remove flicker noise in the image without reducing the image resolution. Then, contrast-limited adaptive image enhancement is used to improve the contrast between the defect and the background in the image.

[0022] In one embodiment of the present invention, the types of defects mentioned in step S4 include: vias, rodent bites, open circuits, short circuits, burrs, and fake copper. The standardized dataset is labeled using the image annotation tool LabelImg. Each labeled PCB image generates a corresponding txt file. The txt file contains the following information: the category information of the target bounding box, the coordinates of the center point, the width, and the height. The target bounding box is a labeled rectangle that surrounds the defect.

[0023] In one embodiment of the present invention, in order to ensure the independence of the dataset in step S4, the dataset is first divided into training and validation sets and test sets in a 9:1 ratio, and then the training and validation sets are divided into training and validation sets in a 9:1 ratio within the training and validation sets.

[0024] In one embodiment of the present invention, the backbone feature network CSPDarknet53 uses a residual network, wherein the residual convolution is divided into two parts: the backbone part is a 1*1 convolution plus a 3*3 convolution; the residual edge part directly combines the input and output of the backbone. This makes the network easy to optimize and can improve accuracy by increasing the depth. In order to further improve the feature extraction capability of the backbone network, a Focus network structure is added to the backbone network. It takes a value for every pixel of the input image to obtain four independent feature layers. Then, the four independent feature layers are stacked. At this time, the width and height information is concentrated in the channel information, thereby expanding the input channels by four times. The concatenated feature layer becomes twelve channels instead of the original three channels.

[0025] In the feature fusion stage, CSPLAYER was reconstructed using ResBlock. The reconstructed CSPLAYER module can better handle multi-scale feature fusion, obtain more semantic information, and alleviate the gradient repetition problem.

[0026] Decoupled Head predicts pixel categories by adding one or more additional branches to the feature map of the backbone network.

[0027] Compared with the prior art, the above-mentioned technical solution of the present invention has the following advantages: The PCB defect detection system of the present invention truly establishes a machine vision system for automatic acquisition and detection of defect images, and the effectiveness of the detection system has been verified on actual PCB boards; the present invention improves the YOLOX algorithm by optimizing data augmentation, reducing the instability caused by single Mosaic data augmentation, and improving the accuracy of the model; the present invention adds a Focus structure to the backbone feature network, effectively expanding the channel information; the present invention also introduces a Decoupled Head as one of the YOLO Heads, improving the detection capability of small targets and better handling the image features of subtle defects. Attached Figure Description

[0028] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0029] Figure 1 This is a schematic diagram of the image acquisition and processing module in the PCB defect detection system based on the improved YOLO algorithm of this invention.

[0030] Figure 2 This is a flowchart of the PCB defect detection method based on the improved YOLO algorithm of this invention.

[0031] Figure 3 This is the original sample image of the PCB dataset described in this invention.

[0032] Figure 4 This is the improved YOLOX algorithm structure described in this invention. Figure I ;

[0033] Figure 4-1 This is the improved YOLOX algorithm structure described in this invention. Figure II .

[0034] Figure 5 This is a diagram showing the defect detection results described in this invention.

[0035] As shown in the figure, 1. Camera; 2. Lens; 3. Clamping device; 4. Ring light source; 5. Computer hardware; 6. Base; 7. Rectangular light source; 8. Light source controller; 9. PCB sample. Detailed Implementation

[0036] like Figure 1As shown, this embodiment provides a PCB defect detection system based on an improved YOLO algorithm. The detection system adopts an improved YOLOX algorithm, which reduces the instability caused by single Mosaic data augmentation by optimizing data augmentation and improving the accuracy of the model. It is characterized by including: an image acquisition and processing module, a neural network model improvement module, a model training module, a model verification module, and a defect detection module.

[0037] The image acquisition and processing module acquires images of PCB surface defects by building an image acquisition system and processes the acquired images to obtain a dataset. The image acquisition and processing module includes a camera, a lens, a clamping device, and a base. A ring light source is also provided on the lens side. After the camera and lens are assembled, they are placed together on the clamping device. The camera's output interface is connected to the computer hardware.

[0038] The neural network model improvement module, model training module, model validation module, and defect detection module all run on computer hardware as data modules. The neural network model improvement module is based on the YOLOX neural network and is used to build a defect detection model. The improvements include using a focus structure to stack feature layers, expanding the input channels; reconstructing CSPLAYer with ResBlock in the feature fusion stage, the reconstructed CSPLAY module can better handle multi-scale feature fusion, obtain more semantic information, and alleviate the gradient repetition problem; then, spatial pyramid pooling is used to process max pooling at four different scales, which greatly improves the receptive field; finally, a decoupled head is used as the YOLO head.

[0039] The model training module is used to set the network parameters of the defect detection model. The created dataset is input into the defect detection model for training. Adam is used as the optimizer during training. The learning rate of Adam optimizer is set to 0.001, the weight decay is set to 0.0005, the number of training epochs is set to 500, and the batch size is set to 4.

[0040] The model validation module is used to test the initially trained defect detection model with a validation set, and adjust the learning rate and epoch parameters according to the validation results to improve the validation accuracy and obtain the final defect detection model.

[0041] The defect detection module is used to input test set data into the trained defect detection model, detect and label the type, location and confidence level of PCB surface defects in the test set.

[0042] The lens ring on the lens and the PCB sample to be tested on the base are on the same vertical straight line. The base is also equipped with a rectangular light source and a light source controller. The input terminal of the light source controller is connected to the computer hardware by a control line. The camera is an industrial CCD camera. A ring light source is coaxially and vertically fixed above the PCB board to be tested. A rectangular light source is placed below the PCB board. The defect image of the PCB surface is obtained by adjusting the distance between the camera, the light source and the PCB board and the exposure time.

[0043] like Figure 2 The flowchart of the PCB defect detection method is shown. This embodiment provides a PCB defect detection method based on the improved YOLO algorithm, including the following steps: building a PCB defect image acquisition system, acquiring PCB defect images and performing image processing and enhancement, creating a dataset, building a training model, training the model and generating a weight file, reading the weight file and testing the test set.

[0044] Step 1: Set up a PCB defect image acquisition system, such as... Figure 1 As shown, the image acquisition and processing module acquires images of PCB surface defects by building an image acquisition system. The camera is perpendicular to the PCB sample, and the image is made clear by adjusting the distance between the camera and the sample.

[0045] Step 2: Acquire PCB defect images and perform image processing and enhancement. The images acquired in Step 1 are standardized. In this embodiment, there are many types of defects in the PCB manufacturing process. The main defect types are: vias, rodent bites, open circuits, short circuits, burrs, and fake copper. The acquired defect images are standardized to 640*640 pixels. Image enhancement is performed on the standardized images. First, the acquired defect images are translated, flipped, and rotated to expand the data. Then, image noise reduction and enhancement are performed on the expanded images. First, a combination of average filtering and smooth filtering is used for noise reduction. Moving is selected as the smoothing method for smoothing, with a span of 5. The window size for average filtering is set to 4. This noise reduction method can effectively remove flicker noise in the image without reducing the image resolution. Then, contrast-limited adaptive image enhancement is used to improve the contrast between defects and the background in the image.

[0046] Step 3: Dataset Creation. The processed image dataset is first divided into training and validation sets and test sets in a 9:1 ratio, and then the training and validation sets are further divided in a 9:1 ratio. In this example, there are 500 images of each type of defect in the dataset, for a total of 3000 images: 2430 images for the training set, 270 images for the validation set, and 300 images for the test set. The image annotation software Labelimg is used to annotate the images in the dataset with defects. Holes are named missing_hole, mouse bites are named mouse_bite, open circuits are named open_circuit, short circuits are named short, burrs are named spur, and fake copper is named spurious_copper. After each image is annotated, Labelimg generates a corresponding txt file containing the category information of the target bounding box and the coordinates (x, y), width, and height (w, h) of the target bounding box's center point.

[0047] Step 4: Model Training and Construction. The YOLO neural network model is built using a combination of Mosaic and Mixup methods for data augmentation. Mosaic first flips and scales the input image, then arranges and stitches the images to ensure the new image contains the previously labeled information. Mixup performs image enhancement by combining images of different classes, thereby expanding the dataset. In the original YOLOX model, data augmentation is turned off in the last 15 epochs to allow the model to converge. However, in our implementation, we found that turning off data augmentation in the last 50 epochs actually resulted in higher model accuracy and improved detection performance. Figure 4The improved YOLOX network structure mainly includes a CSPDarknet53 backbone feature extraction network with a Focus structure, a Spatial Pyramid Pooling (SPP) structure, a Path Aggregation Network (PAFPN), and a YOLO Head with a Decoupled Head. The improved backbone feature extraction network CSPDarknet53 incorporates a Focus structure. The Focus structure, first proposed in YOLOv5, segments the original input image, then concatenates the segments, and finally performs a convolution to reduce the original input image size from 640*640*3 to a 320*320*64 feature image. This structure reduces the loss during image downsampling. The CSPDarknet53 model, in its original YOLOX form, primarily consists of CBS (convolutional layers + normalization layers + activation functions) and CSPrayer. The CSPrayer layer itself is also composed of CBS and residual blocks (CBS). This approach allows for deeper neural networks, thereby improving feature extraction and fusion capabilities. However, in practical applications, it has been found that this model's learning and representational abilities are weak when dealing with small target defects. Therefore, this invention reconstructs the CSPrayer using Res2Block. Res2Block can better handle multi-scale feature fusion, acquire more semantic information, and alleviate the gradient repetition problem. Figure 4 CSPRB_X represents the reconstructed CSPLAYer structure. The SPP structure performs convolution on the features obtained from Dark4 followed by max pooling. Feature extraction is achieved through max pooling with different kernel sizes, significantly increasing the receptive field. The head detection head in this invention employs a decoupled head structure, which preserves more detailed features of the target. These operations can obtain more detailed features for minor defects in PCBs, effectively enhancing the small target feature extraction and detection capabilities of the feature extraction network.

[0048] Step 5: Train the model and generate a weight file. Input the training set into the constructed training network structure model for training. First, set the initial network parameters for the defect detection model. Adam is used as the optimizer during training. The learning rate of the Adam optimizer is set to 0.001, the weight decay is set to 0.0005, the number of training epochs is set to 500, and the batch size is set to 4. In the specific training process, the network first divides the input image into grids of different sizes to segment it. Each grid is independently responsible for a different region. If the center of the target to be detected is in a certain grid, then that grid is responsible for detecting the target.

[0049] Step 6: Validate the pre-trained defect detection model using validation set images. Adjust the network parameters based on the validation results, including learning rate, batch size, and weight decay. After modification, retrain until the validation results meet expectations.

[0050] Step 7: Read the weight file and test the test set. Use a Python script to read the weight file and input the prepared test set into the trained model for testing. The generated detection result images will be saved in a separate folder. Once all images in the test set have been tested, a "finish" message will be displayed. The average detection time per image is approximately 0.7 seconds. Test results are as follows: Figure 5 As shown.

[0051] When the PCB defect detection method proposed in this embodiment is applied to industrial production, the PCB image acquired according to the method in step one is input into the trained improved YOLOX training network structure model, which can realize the real-time identification, classification and localization of surface defects. This plays a huge role in helping technicians troubleshoot machine faults and improve production quality.

[0052] Single-object detection algorithms based on the YOLO series achieve high detection speeds while sacrificing a small amount of accuracy, and are widely used in industrial scenarios. The improved YOLOX algorithm used in this patent, compared to other YOLO series algorithms, adopts an anchor-free strategy. It no longer relies on predefined anchor boxes, but adaptively adjusts by predicting the object's center point and size, thus improving detection flexibility. Mosaic and Mixup are used to augment the input data, effectively increasing data diversity and improving the model's generalization ability. In the feature fusion stage, CSPLAYER is reconstructed using ResBlock. The reconstructed CSPRB module can better handle multi-scale feature fusion, acquire more semantic information, and alleviate gradient repetition problems. SPP is added to the backbone network to increase the network's receptive field. PANet is used to better fuse multi-layer features, improving the model's detection ability in situations with small objects and complex backgrounds. A decoupling head is used to improve prediction accuracy.

[0053] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A PCB defect detection system based on an improved YOLO algorithm, wherein the detection system employs an improved YOLO algorithm, which reduces the instability caused by single-mosaic data augmentation by optimizing data augmentation, thereby improving the accuracy of the model, characterized in that... include: Image acquisition and processing module, neural network model improvement module, model training module, model verification module, defect detection module; The image acquisition and processing module acquires images of PCB surface defects by building an image acquisition system and processes the acquired images to obtain a dataset. The image acquisition and processing module includes a camera (1), a lens (2), a clamping device (3), and a base (6). At the same time, a ring light source (4) is provided on the side of the lens (2). After the camera (1) and lens (2) are assembled, they are placed together on the clamping device (3). The output interface of the camera (1) is connected to the computer hardware (5). The neural network model improvement module, model training module, model verification module, and defect detection module are all running on computer hardware (5) as data modules. The neural network model improvement module is based on the YOLOX neural network and is used to improve and construct a defect detection model. The improvement includes using the Focus structure to stack the feature layers and expand the input channels. In the feature fusion stage, CSPLAYER was reconstructed using ResBlock. The reconstructed CSPLAYER module can better handle multi-scale feature fusion, obtain more semantic information, and alleviate the gradient repetition problem. Then, spatial pyramid pooling was used to process max pooling at four different scales, which greatly improved the receptive field. Finally, a decoupled head was used as the YOLO head. The model training module is used to set the network parameters of the defect detection model. The created dataset is input into the defect detection model for training. Adam is used as the optimizer during training. The learning rate of Adam optimizer is set to 0.001, the weight decay is set to 0.0005, the number of training epochs is set to 500, and the batch size is set to 4. The model validation module is used to test the initially trained defect detection model with a validation set, and adjust the learning rate and epoch parameters according to the validation results to improve the validation accuracy and obtain the final defect detection model. The defect detection module is used to input test set data into the trained defect detection model, detect and label the type, location and confidence level of PCB surface defects in the test set.

2. The PCB defect detection system according to claim 1, characterized in that: The light-collecting lens ring on the lens (2) and the PCB sample (9) to be tested on the base (6) are on the same vertical straight line. The base (6) is also provided with a rectangular light source (7) and a light source controller (8). At the same time, the input terminal of the light source controller (8) is connected to the computer hardware (5) by a control line.

3. A PCB defect detection method based on an improved YOLO algorithm, constructed based on the PCB defect detection system described in claim 1, characterized in that, Includes the following steps: Step S1: Set up an image acquisition system to acquire images of defective samples; Step S2: Edit the acquired defect images to obtain images with uniform dimensions; Step S3: Use data augmentation methods to augment the samples, and then perform noise reduction and contrast enhancement on the images; Step S4: Select image samples from the images, label the location and type of defects in the image samples, and create a dataset; Step S5: Input the processed RGB three-channel image into the improved YOLOX neural network for training to obtain a target detection model for PCB panel defects; Step S6: Read the obtained target detection model and test it on the test set prepared in step S4 to realize the identification, classification and location of PCB defects; The target detection model in specific step S5 includes: a CSPDarknet53 backbone feature extraction network with a focus structure, a spatial pyramid pooling structure SPP, a path aggregation network PAPPN, and a YOLOHead with a decoupled head.

4. The PCB defect detection method according to claim 3, characterized in that: The size of the image after unification in step S2 is 640*640.

5. The PCB defect detection method according to claim 3, characterized in that: The data augmentation method described in step S3 is as follows: First, the acquired defect image is translated, flipped, and rotated to augment the data. Then, the augmented image is subjected to image denoising and enhancement. First, the image is denoised by combining average filtering and smooth filtering. Moving is selected as the smoothing method for smoothing, and the span is set to 5. The window number for average filtering is set to 4.

6. The PCB defect detection method according to claim 3, characterized in that: The types of defects mentioned in step S4 include: vias, rodent bites, open circuits, short circuits, burrs, and fake copper. The standardized dataset is labeled using the image annotation tool LabelImg. Each labeled PCB image generates a corresponding txt file. The txt file contains the following information: the category information of the target bounding box, the coordinates of the center point, the width, and the height. The target bounding box is a labeled rectangle that encloses the defect.

7. The PCB defect detection method according to claim 3, characterized in that: In step S4, to ensure the independence of the dataset, the dataset is first divided into training and validation sets and test sets in a 9:1 ratio, and then the training and validation sets are divided into training and validation sets in a 9:1 ratio.

8. The PCB defect detection method according to claim 3, characterized in that: The backbone feature network CSPDarknet53 uses a residual network, where the residual convolution is divided into two parts: the backbone part consists of a 1*1 convolution followed by a 3*3 convolution; the residual edge part directly combines the input and output of the backbone. This makes the network easy to optimize and allows for increased accuracy by increasing depth. To further improve the feature extraction capability of the backbone network, a focus network structure is added. This structure takes a value from every pixel of the input image, resulting in four independent feature layers. These four independent feature layers are then stacked, concentrating the width and height information into the channel information, thus expanding the input channels fourfold. The stacked feature layer becomes twelve channels instead of the original three. In the feature fusion stage, CSPLAYER was reconstructed using ResBlock. The reconstructed CSPLAYER module can better handle multi-scale feature fusion, obtain more semantic information, and alleviate the gradient repetition problem. Decoupled Head predicts pixel categories by adding one or more additional branches to the feature map of the backbone network.

Images