Research On Object Detection Methods Of Printed Circuit Board Assembly Scene Based On The Convolutional Neural Network

The electronic product manufacturing industry is a strategic,fundamental,and leading pillar industry of the national economy.The technical requirements for high density and efficient and precise assembly of circuit board electronic components have prompted the intelligent transformation and upgrading of visual object detection technology in the assembly process of electronic products.However,for the current machine vision algorithms based on convolutional neural networks,due to the lack of in-depth research on "weak feature representation of small-sized objects,the internal interpretability mechanism of convolutional neural networks,and the problem of visual scene imbalance," it is difficult to exert its advantages under the application requirements of high accuracy,lightweight and fast object detection model in the circuit board assembly scene,and it cannot guarantee the quality and efficiency of the circuit board assembly process,thereby affecting the cost,reliability,and time-to-market of electronic products.In this context,this paper takes "the study of object detection method in circuit board assembly scene based on convolutional neural network" as the starting point,and takes the object in the surface mount/mixed assembly scene and the through-hole assembly scene involving the circuit board assembly process as the detection object.Starting from the data level of convolutional neural network,backbone network,feature fusion strategy and detection head,it provides an important theoretical and technical reference for promoting the development of circuit board assembly in the direction of intelligence,precision and agility.The specific research contents and conclusions are as follows:(1)A small object detection method for PCB electronic components based on multiple detection heads is proposed: Aiming at the problem of low detection accuracy caused by the high density and small size of electronic components on the circuit board,it is found that the low pixel occupancy of the small object leads to its weak feature representation.According to the principle that the shallow detection head can effectively express the weak feature of small objects,an object size-backbone feature preservation correspondence quantization method is proposed,which increases detection heads,multi-dimensional anchors,and feature fusion paths sensitive to smallsized objects by weighing object size and computing power.The method test experiment was completed in the self-built electronic components joint data set OPCBA-29(Objects in Printed Circuit Board Assembly,OPCBA).The results show that the object size-backbone network feature retention correspondence quantification method can effectively mine the weak features of small objects in the backbone network.Compared with the original benchmark model,the m AP(Mean Average Precision)of29 classes of electronic component objects increased from 77.08% to 93.07% based on the multi-detection head PCB electronic component small object detection method.(2)A lightweight detection method for PCB electronic components based on effective receptive field-anchor matching is proposed: Aiming at the problems of redundant structure and large amount of model parameters based on the backbone network model of convolutional neural network.According to the principle of explaining that the internal reasoning decision of the open convolutional neural network can bring high reliability and lightweight network,the calculation and visualization method of the effective receptive field size of different depth layers of convolutional neural network based on gradient backpropagation is designed,and a modular deconstruction combination method of backbone network based on effective receptive field-anchor matching strategy is proposed,and the module reconstruction of the backbone network is realized by using this strategy.Through the self-built dataset OPCBA-29* to complete the method test experiment,the experimental results show that the object detection model of PCB electronic components based on effective receptive field-anchor matching,compared with the original benchmark model,the m AP of 29 classes of object detection reaches 95.03%.The model parameters are only35.61% of the original parameters.It is verified that the modular reconstruction of the backbone network based on the effective receptive field research of the internal visual mechanism can improve the detection accuracy and realize the model’s lightweight.(3)An object detection method based on effective receptive field-anchor allocation in PCB through-hole assembly scene is proposed: Because of the interclass high similarity and intra-class large differences objects to be detected in the PCB through-hole assembly scene,which seriously affect the detection accuracy.According to the principle that the precise identification and positioning analysis inside the convolutional neural network can effectively mine the separable features between classes and the compact features within the class,a refined analysis of the grid’s effective receptive field on the detection head was carried out,and the influence mechanism of the detection head grid on the object detection effect when anchors of different sizes were placed corresponding to the effective receptive field was studied.The effective receptive field range based on the detection head was proposed,and the precise anchor allocation method based on the proposed method can effectively discover inter-class separable features and intra-class compact features.Through the self-built dataset OPCBA-21 to complete the experiment,the m AP increased from79.32% to 89.86% based on the effective receptive field-anchor allocation method.It is verified that the precise anchor allocation based on the refined analysis of the detection head’s effective receptive field effectively improves the detection accuracy of objects with high inter-class similarity and large intra-class differences.(4)A fast and accurate object detection method for PCB through-hole assembly scene based on balance strategy is proposed: Aiming at the problems of sample class imbalance,object scale imbalance,and positive/negative sample imbalance that affects the speed and accuracy of object detection in PCB through-hole assembly scenarios.According to the balance strategy,the detection accuracy and detection speed can be weighed,a balanced division method of training set/validation set proportion is proposed.An "additive" feature fusion strategy is designed that balances deep object location information and shallow object semantic information.An effective anchor concept and solution method for correcting the imbalance of positive and negative samples and avoiding data redundancy are proposed.The effective anchor completes accurate anchor allocation while removing redundant anchors.The method test experiment is completed using the self-built data set OPCBA-21*.Compared with the original benchmark model,the experimental results show that the object detection method based on the balance strategy in the PCB through-hole assembly scene increased the m AP from 89.78% to 94.25% on 21 classes of objects.The positioning accuracy has increased from 49.54% to 54.20%,and the average forward inference time per test image has dropped from 9.21 ms to 3.22 ms.This method validates the effectiveness of improving accuracy and speed in addressing the object imbalance problem of convolutional neural networks.This paper originates from the engineering requirements of object detection in intelligent circuit board assembly scenarios.It uses convolutional neural networks to address the critical issues involved in " object detection of electronic components in surface mount/mixed assembly scenarios" and " object detection in through-hole assembly scenarios" A series of studies have been carried out to verify the correctness and effectiveness of the proposed theories and methods.Here we use detection accuracy,model parameters and detection speed as evaluation indicators.The research results can be extended to object detection in different fields,laying a foundation for future related visual detection research based on deep learning.