Research And Application Of Circuit Surface Anomaly Detection Technology Based On Deep Learning

With the development of the electronics industry,practitioners and researchers are increasingly concerned about the reliability issues in the production process of printed circuit boards.As the integration level of circuit board continues to increase,defects are more likely to occur during the circuit manufacturing process,and the existence of such defects can directly affect the life and reliability of the equipment.In the production process,detecting defects in the circuit as early as possible can minimize the impact of defects on the equipment,thereby reducing equipment maintenance costs.Therefore,how to detect circuit defects during the circuit manufacturing process has important engineering significance and research value.To address this issue,this thesis proposes a deep learning-based anomaly detection technique for circuit surfaces.First,the circuit image is preprocessed for calibration using an image calibration algorithm.Then anomaly detection and defect localization are performed on the circuit image by an anomaly detection model.Finally,the object detection algorithm and model pruning method are used to improve the accuracy of anomaly detection and realize edge-side anomaly detection.The main work of this thesis is as follows:(1)Aiming at the problem that the traditional ORB algorithm does not have scale invariance and the accuracy of feature point matching is low,an image calibration algorithm based on Improved Oriented Fast and Rotated Brief(IORB)is proposed.The algorithm detects key points by constructing a scale space,and generates descriptors for the detected key points through the r BRIEF method,then calculates the Hamming distance for the obtained feature points for rough matching,and finally adopts RANSAC algorithm removes the wrong matching points to obtain the final matching result.Experimental results show that the IORB algorithm achieves high matching accuracy while obtaining scale invariance.(2)In response to the insufficient extraction of differential feature information from images and the underutilization of shallow network representation information in traditional Patch SVDD,this thesis introduces the Cut-Paste method to randomly crop and paste image patches,and trains an encoder and a classifier to encode and classify the image patches,thereby enhancing the ability of the encoder to extract differential features between normal and abnormal image patches.At the same time,the shallow and deep information of the encoder is aggregated to help the neural network better utilize image detail information when reading deep semantic information.Experimental results show that the improved Patch SVDD algorithm increases the accuracy by 5%compared to the original method.(3)To improve the accuracy of anomaly detection models in the presence of multiple circuit types and enable anomaly detection in low-computational-resource environments,an anomaly detection system based on Jetson Nano was designed.The system utilizes the YOLOv5(You Look Only Once Version 5)object detection model to detect circuit types and loads the corresponding anomaly detection model based on the detection results.This approach makes anomaly detection models work more efficiently.Furthermore,the system reduces model parameter and computational complexity through sparse training and model pruning methods,enabling edgebased circuit anomaly detection.