Research On PCB Defect Detection Method Based On Deep Learning

With the rapid development of machine vision and intelligent inspection technology,Automated Optical Inspection(AOI)is widely used for surface defect detection.This paper aims to detect the defects such as missing-solder,missing-prteus,overturned-direction,solder-ball and solder-short,etc.,which occur in the soldering process of PCB mounted components,and develops a PCB defect detection system.The main contents of the thesis research include:1.This paper analyzes the inspection tasks,performance requirements and functional modules of the PCB mounting defects inspection system.First,this paper selects appropriate light sources,cameras,mechanical structure and other hardware equipment to construct a good imaging system;Secondly,analyze the PCB defect types and the process of labeling,and perform statistical analysis on the characteristics of image data,provide guidance for the generation of target anchor frames,and clarify the research focus of this article.2.This paper studies an improved free anchor matching on PCB defect detection algorithm.First of all,to detect small PCB defects and poor use of traditional detection methods,this paper uses the K-Means clustering to generate the anchor shapes and box sizes so that they is suitable for PCB defects.Meanwhile,this paper introduces a balanced pyramid network to solve the problem of feature pyramid network semantic information diluting small target features,in which the output features of the feature pyramid network are scaled,integrated and refined,and Gaussian Non-local Attention mechanism is introduced to strengthen the global information of the balanced features,thereby this operation can strengthen the model's attention to the location information of small targets.Secondly,this paper discards the Io U strategy of divising the anchor,and introduces a free anchor matching mechanism based on maximum likelihood estimation to achieve the perfect matching between the anchor and the defect features,which can solve the problem that a single anchor may contain multiple defects while the PCB contains dense components.The experimental results show that the proposed method has a good detection rate for small defects and PCBs containing dense components.The detection rate accuracy is 95.10%,and the detection speed is 8.9FPS/s for the test sample with the resolution of 1280×1024.3.This paper proposes an improved Deep Labv3+ algorithm for PCB defect detection.In order to detect small PCB defects and irregular defect shapes,this paper firstly introduces the backbone network Res Ne St which contains the Split-Attention mechanism(Split-Attention Blocks),in which the features are divided and combined along the channel by grouping convolution to get weight distribution to obtain the relevance of features between channels,and strengthen the model's attention to small targets.Secondly,this paper reconstructs the decoding area of Deep Labv3+,and introduces the downsampling coefficients in the feature extraction network which are1/4,1/8,and 1/16 to get the feature map to enrich small targets' information.Finally,the Online Hard Example Mining is introduced to optimize the areas with large losses in the training process.The experimental results show that the proposed method has a good effect on defect sematic segmentation,contour and edge recovery,and the semantic segmentation m Io U rate is 78.43% and.m Acc rate is 89.64%?4.This paper develops a visual inspection system for PCB surface mount defects based on proposed defect detection algorithms.The detection software can perform real-time detection of PCB product and display the results in the window to provide relevant references for workers' operations.