Research And Application Of Object Detection Method Based On Improved YOLOv4

Printed Circuit Board(PCB)is an important part of electronic products,and its quality detection has become a prerequisite for meeting the long-term normal operation of electronic products.Although deep learning object detection methods have made great progress in the field of PCB surface defect detection,there are still problems such as lack of label sample data,limited feature learning capabilities,difficulty in deep network training,and uneven detection accuracy and speed.Therefore,based on the analysis of the YOLOv4 object detection method,this thesis proposes a PCB surface defect detection scheme design based on the improved YOLOv4.The main work of the thesis is as follows:1.This thesis analyzes the current research status and existing problems of existing PCB surface defect detection methods at home and abroad,and proposes a PCB surface defect detection model based on improved YOLOv4.At the same time,this thesis uses the Multistage Residual Hybrid Attention Module(MRHAM)to improve feature learning to enhance the feature expression ability of the shallow network,so that the receptive field pays more attention to defect object features and ignores irrelevant features;This thesis uses the K-means++clustering algorithm to cluster the PCB surface defect dataset to determine the value of the anchor boxes;This thesis uses online and offline data augmentation,transfer learning and multi-scale training methods to enhance the model's adaptability to different input image scales and improve the stability and generalization ability of the model.2.This thesis designs a PCB surface defect detection software based on the improved YOLOv4 model.The software mainly includes database design,data acquisition,data preprocessing,detection model improvement and optimization implementation,and visual interface design.The database is used to store data information such as users,workshops,and detection results.The data preprocessing function is divided into operations such as data filtering and data labeling to obtain the dataset format that can be directly used by the detection model.The improvement and optimization of the detection model is the core of the entire software.This thesis uses two deep learning frameworks for modeling,and proposes a model improvement and optimization method for the deficiencies of the model in PCB surface defect detection.Finally,this thesis uses the back-end server Flask framework to call the detailed information of the detection results output by the API interface of the deep learning Py Torch framework,and realizes the visual display of the online real-time detection results of the four workshops on the Web side through the front-end Vue framework.3.This thesis builds a system test environment to perform functional tests on the four software functional modules of data acquisition,data preprocessing,detection model improvement and optimization implementation,and visual interface design;analyze and improve the YOLOv4 model and other traditional and deep learning object detection models SSD,Retina Net,Faster R-CNN,YOLOv3 and YOLOv4 detection performance comparison.Experimental results show that compared with other traditional and deep learning object detection models,the improved YOLOv4 model has improved detection accuracy and detection speed.Its average accuracy(m AP)value reaches 99.71%,and the detection speed(FPS)value reaches 68 f.s^-1,indicating that the improved YOLOv4 model has achieved the expected purpose in the application of PCB surface defect detection.