Research On Target Semantic Segmentation Method Of Multi-Component Parts In Industrial Scene

With the continuous improvement of industrial automation,the recognition and segmentation of the target semantic categories of multi-component parts in industrial scenes is the key technology in the field of industrial visual automation detection.This paper aims at the problems of low accuracy of target semantic classification recognition and segmentation positioning of multi-component parts,insufficient target feature information extraction and inaccurate target edge segmentation by existing methods in industrial scenarios.Research multi-scale feature extraction and cascading multicomponent parts target semantic segmentation method.The main research contents of this article are as follows:(1)The current advanced target segmentation algorithm is analyzed.According to the advantages of this kind of algorithm,the residual network model combined with atrous convolution is established.This model is defined as the basic model of this paper.A total of ten types of component parts data sets for bearings,screwdrivers,gears,pliers,wrenches,bolts,nuts,washers,hammers,and files under industrial scenarios were produced.The experimental results show that,under this basic model,the mean intersection over union accuracy of the ten types of component parts is 61.21% and the mean pixel accuracy is 85.41%.It is found that due to the objective complexity and high dynamic range in industrial scenes,the existing segmentation methods have a lot of room for improving the accuracy of the semantic segmentation of multi-component parts targets.(2)In view of the limitation of single-scale feature extraction encoding-decoding network,it is impossible to effectively extract the category information of each target in the multi-object segmentation task and the position information of the corresponding category on the original image.A multi-scale feature extraction module is added to the decoding part,and the feature map output from the encoding part is used to extract features of different scales through parallel artous convolution,which enriches the category information and position information of each target in the feature map output by the decoder.The experimental results show that,compared with the basic network model,the mean intersection over union accuracy of the ten types of component parts is improved by 2.57%,and the mean pixel accuracy is improved by 4.84%.(3)Aiming at the problem that the encoding-decoding network model segments the background into the target wrongly,and the target edge segmentation is not accurate enough,a cascading parts semantic segmentation method is proposed.By converting multi-target segmentation tasks into multiple single-target segmentation tasks,the proportion of segmented background is reduced,the probability of the model segment the background into targets is reduced,the complexity of feature extraction is reduced,and the target edge segmentation is more accurate.The experimental results show that,compared with the basic network model,the mean intersection over union accuracy is improved by 12.89% of the cascading semantic segmentation method,the mean pixel accuracy is improved by 8.57% of the cascading semantic segmentation method.In this paper,a cascading semantic segmentation method is proposed,which improves the ability of the encoding-decoding model to extract target features of multiple types of parts and the edge segmentation accuracy of the target,and reduces the probability of wrong segmentation of the model.It provides a method basis and reference value for the accurate identification and positioning of multi-component parts targets in complex industrial scenarios.