Method for establishing modified-yolo-based surface defect detection method and defect-detection system

Abstract

A method for establishing a surface defect detection model based on a modified YOLO model and a defect detection system running the method are disclosed. The defect detection system acquires images of the object to be detected using a photographic module, preprocesses and annotates the images to form a surface defect image dataset, introduces a higher resolution feature layer and adopts a bidirectional feature fusion strategy, combines the attention mechanism of a simple attention module and a multi-scale feature extraction module to realize a multi-scale simple attention module, so as to accurately extract the key features of small target defects of the object to be detected. Then, the surface defect image dataset is used to train a surface defect detection model to perform small target defect detection on the object to be detected, and outputs the surface defect detection results.

Claims

1. A method for establishing a surface defect detection model based on a modified YOLO model, running in a defect detection system equipped with the modified YOLO model, the method comprising: The process involves: acquiring an image of an object to be detected; preprocessing and labeling the image to form a surface defect image dataset; introducing a higher-resolution feature layer, specifically a small-object detection layer that dynamically upsamples to generate a high-resolution feature map for defect detection, progressively upsampling depth-scale features to shallow-scale high-resolution features to enhance the surface defect detection model's ability to detect small-object defects on the object; combining an attention mechanism in a convolutional neural network, using 7×7, 11×11, or 23×23 convolutional kernels and horizontal and vertical convolutions to capture small, medium, and large-sized image features to accurately extract key features of small-object defects on the object; and using this surface defect image dataset, training a machine learning algorithm to perform small-object defect detection on the object, outputting a surface defect detection result.

2. The method for establishing a surface defect detection model based on a modified YOLO model as described in claim 1, wherein when acquiring an image of the object to be inspected, an 8K line scan camera and a 3D line scanner are simultaneously activated to acquire high-resolution planar images and stereo point cloud data of the object to be inspected.

3. The method for establishing a surface defect detection model based on a modified YOLO model as described in claim 1, wherein, The surface defect image dataset is obtained, and a data augmentation is performed to expand the surface defect image dataset, and a training set, a validation set and a test set are established.

4. The method for establishing a surface defect detection model based on a modified YOLO model as described in claim 1, wherein, In the step of extracting key features of small target defects of the object to be detected, a MobileNet general message bottleneck C2f module is used to improve the extraction efficiency of key features of small target defects by using depthwise separable convolution and point convolution design.

5. The method for establishing a surface defect detection model based on a modified YOLO model as described in claim 4, wherein the MobileNet general message bottleneck C2f module introduces a lightweight depthwise separable convolution and a pointwise convolution, decomposing the MU bottleneck portion of a complete convolution operation into an initial depthwise convolution, a dilated convolution, an intermediate depthwise convolution, and a projective convolution.

6. A method for establishing a surface defect detection model based on a modified YOLO model as described in any one of claims 1 to 5, wherein the object to be detected is a high-density interconnect printed circuit board, and the small target defects include at least circuit board warpage, solder joint height, and microstructural defects.

7. A defect detection system, wherein a modified YOLO model is provided, the defect detection system comprising: A surface defect detection platform forms a processing plane and is equipped with a computer, a signal processing module, a photographic module and a light source module; The method for establishing a surface defect detection model based on the modified YOLO model using the computer includes: acquiring an image of an object to be detected using the photographic module; preprocessing and labeling the image of the object to be detected to form a surface defect image dataset; introducing higher-resolution features, wherein a higher-resolution small target detection layer is introduced to generate a high-resolution feature map for defect detection through dynamic upsampling, progressively upsampling depth-scale features to shallow-scale high-resolution features to enhance the surface defect detection model's ability to detect small target defects of the object to be detected; combining an attention mechanism in a convolutional neural network, using 7×7, 11×11, or 23×23 size convolutional kernels and horizontal and vertical convolutions to capture small, medium, and large image features to accurately extract key features of small target defects of the object to be detected; and using the surface defect image dataset, training the surface defect detection model for small target defect detection of the object to be detected using a machine learning algorithm, performing small target defect detection on the object to be detected, and outputting a surface defect detection result.

8. The defect detection system as described in claim 7, wherein the photographic module includes a 3D line scanner and an 8K line scanner camera, and is electrically connected to the computer and the signal processing module via a data cable, and the 8K line scanner and the 3D line scanner are synchronously started to acquire high-resolution planar images and stereo point cloud data of the object to be inspected.

9. The defect detection system as described in claim 7, wherein the light source module is provided with a light source and a light source controller, the light source is a multi-band light source, and when the camera module is started to acquire the image of the object to be inspected, the light source controller synchronously controls the light source to flash and drives the light source to emit multi-band light to illuminate the object to be inspected on the processing plane.

10. The defect detection system as described in claim 9, wherein the multi-band light source integrates ultraviolet light, visible light and infrared light, and dynamically switches between different bands of the multi-band light source to detect various types of defects on the object to be detected.

11. The defect detection system as described in claim 9, wherein the light source module incorporates a polarizer and a polarizing filter, and the polarizer adjusts a polarization angle to reduce light interference to the highly reflective areas of the object to be inspected.

12. The defect detection system as described in claim 7, wherein the surface defect detection platform is provided with a color mark sensor for detecting the incoming material condition of the object to be inspected.

13. The defect detection system as described in claim 7, wherein, In the step of extracting key features of small target defects of the object to be detected, a MobileNet general message bottleneck C2f module is used to improve the extraction efficiency of key features of small target defects by using depthwise separable convolution and point convolution design.

14. The defect detection system as described in any one of claims 7 to 13, wherein the object to be detected is a high-density interconnect printed circuit board, and the small target defect includes at least circuit board warpage, solder joint height and microstructural defects.

Images