Design Of Embedded Image Processing System Based On Printing Defect Detection

With the development of modern society,digitalization and information technology has become a major trend in the development of science and technology,electronic products are widely used in unmanned,consumer electronics,medical equipment,aerospace,and many other fields,known as the "mother of electronic products" printed circuit board(Printed Circuit Board,PCB)is an important part of various types of electronic products.is an important part of various types of electronic products.To improve the quality of electronic products and extend the life of electronic products,many scholars at home and abroad are committed to the research of printed circuit board defect detection.Through printed circuit board defect detection,the problem of unstable electronic product quality can be effectively solved at the hardware level,thus guaranteeing the reliable operation of electronic products.This thesis focuses on establishing a lightweight neural network model for PCB defect detection and designing a hardware-based inference framework for the model,and finally building an embedded defect detection system for testing,as follows:(1)Establishing a defect detection model: In this paper,based on the analysis of the characteristics of PCB defects and the calculation principle of inference of convolutional neural networks,YOLOV4-Tiny is used as the defect detection algorithm to establish an algorithmic model for PCB defect detection.The average value of accuracy evaluated by the test set is 91.9%,which is more balanced in terms of accuracy and speed,and the detection performance meets the needs of PCB defect detection in actual production.(2)Hardware inference implementation: Based on ZYNQ hardware characteristics and YOLOV4-Tiny network structure,we design a software and hardware collaborative yolov4-tiny image processing module.In this paper,the PS side is responsible for process control and read/write control in hardware inference,and the computationally intensive and highly parallelized convolution,activation,and upsampling computations are completed in the PL side.In order to reduce the computational cost and improve the operational efficiency,the BN layer is fused with layer parameters and the weight parameters are fixed-point operated;and the data transfer method and the cyclic parallel unfolding and pipeline processing are optimized to reduce the computational latency.(3)Embedded system building: This paper builds a complete embedded PCB defect detection system based on Xilinx ZYNQ hardware platform for image acquisition,processing and result display.Based on the modular design idea,this paper designs the image acquisition module IP core,the result output display module IP core,builds a complete hardware system based on AXI interconnection module,and also designs the algorithm control flow to schedule each functional module to achieve the goal of PCB defect detection in the embedded system.Finally,the designed embedded defect detection image processing system is experimentally compared with the defect detection technology under CPU and GPU platforms.The performance of the embedded detection system designed in this paper is lower than GPU,existential advantages over CPU performance,and there are also better results compared with other embedded platforms.