| Printed Circuit Board(PCB)component detection algorithms,as the core technology of automatic optical inspection systems,have been widely used in enterprise PCB quality inspection,malicious hardware Trojan horse detection,PCB reverse engineering,and other scenarios,with tremendous research and application value.However,the current mainstream PCB component detection algorithms have the following shortcomings:(1)Insufficient use of visual features of PCB images designed by humans;(2)Ignoring the layout structure information of components that can help improve detection accuracy.To address these issues,this thesis focuses on two key issues of PCB visual feature fusion and PCB component relationship inference,and conducts in-depth research on PCB component detection tasks,combining graph neural networks and object detection-related theories.The main research contents of this thesis are as follows:(1)A PCB component detection algorithm based on a multi-level feature fusion graph neural network is proposed.It can efficiently fuse three visual features of PCB images and use the fused features for PCB component detection.First,to construct visual feature images into nodes in a graph neural network,this thesis uses a feature extraction backbone network based on a graph neural network to extract multiple different scale features of visual feature images and construct them into multi-scale feature nodes.Second,to efficiently fuse multi-scale feature nodes constructed from three visual features,this thesis uses a graph neural network to model the complex relationship between different feature nodes,thereby avoiding the problem of interference between different features caused by direct feature fusion.(2)A PCB component recognition algorithm based on a relationship inference graph neural network is proposed,which can learn the relationship between components from the layout structure information and use this relationship for joint inference of all component categories on the PCB.First,all components on a PCB are constructed into nodes of a graph,and then the layout structure information of the components is transformed into the structure information of the graph by connecting nodes that are close in position using the K-nearest neighbor algorithm.Second,by using the node message passing mechanism of the graph neural network,each node can learn the feature of adjacent nodes in its neighborhood and local graph structure information.Finally,each node uses the learned neighborhood information to perform joint inference on all component categories.(3)For the algorithms proposed in the above work,this thesis conducted extensive experiments on the largest PCB component detection dataset FICS-PCB and compared them with other state-of-the-art algorithms.The experimental results show that the average precision(AP)of the proposed algorithm for detection is 0.651,and the average accuracy(Acc)for classification is 89.85%,both of which exceed the current known stateof-the-art algorithms,verifying the effectiveness of the proposed algorithm.This thesis innovatively utilizes the advantages of graph neural networks in feature fusion and relationship inference to solve the two problems existing in current PCB component detection algorithms.The proposed algorithm not only exceeds the current stateof-the-art algorithms but also is the first algorithm to apply graph neural networks to PCB component detection tasks. |
