Research On Part Object Detection And Semantic Segmentation Based On Deep Learning

Nowadays,in the field of industrial production,with the introduction of deep learning technology,intelligent manufacturing has become the mainstream direction of industrial production,and is also synonymous with modern factories.Intelligent manufacturing has high requirements for intelligent assembly technology,because it needs to pay attention to the efficiency of product production and product quality.Object detection and semantic segmentation based on deep learning are the key technologies of intelligent assembly and are the research focus of scholars.The traditional algorithms of object detection and semantic segmentation rely on manual operation in the stage of feature extraction,which leads to the slow speed and low precision of the traditional algorithms.In the actual production and application,there will be interference factors such as complex and diverse environments of parts,similar shapes and sizes of parts,parts shielding and so on,which will lead to low accuracy of parts detection,high missing rate,easy loss of parts segmentation pixels and so on.These problems are difficult to be effectively solved by traditional algorithms,let alone meet the needs of actual production.Therefore,this paper takes parts as the research object and improves the YOLOv4 detection network and Deeplabv3+segmentation network respectively based on the research and analysis of object detection and semantic segmentation in the field of deep learning.The improved network was validated with a public inspection and segmentation dataset,PASCAL VOC 2007,and a homemade part detection and segmentation dataset.The main research work of this paper is as follows:(1)Aiming at the problem that there are few data sets of existing public parts,this paper selects four different parts in terms of size,material,color,etc.,and makes corresponding data sets of object detection and semantic segmentation of parts through data acquisition,data annotation,data set design,etc.,so as to be used in subsequent experiments.(2)Aiming at the YOLOv4 target detection network,the precision of parts detection is not good and there are some problems such as missing detection.In this paper,the network structure is improved and optimized.Firstly,two new SPPNet structures are added between PANet and YOLO Head in the YOLOv4 network structure.Secondly,the kernel parameters of SPPNet are adjusted and optimized to make the improved network have better detection effect.This paper verifies the universality of the improved network through a common data set,PASCAL VOC 2007,and then verifies that the improved network is superior to the original YOLOv4 network for parts detection by using a homemade part target detection data set.(3)Aiming at the problems such as pixel loss and incomplete contour edge segmentation when Deeplabv3+ semantic segmentation network is used for parts segmentation,this paper makes two improvements on it.Firstly,coordinate attention mechanism(CA)is introduced between the backbone feature extraction network MobileNetv2 and the empty space convolution pooling pyramid(ASPP)and after the ASPP module to improve the extraction of pixel position information.Secondly,feature fusion is carried out on the four convolutional feature layers in ASPP module to improve the extraction of deep feature information.The experimental part includes a public data set PASCAL VOC 2007 segmentation data set to verify the ubiquity of the improved network and a homemade part segmentation data set to verify the effectiveness of the improved Deeplabv3+network for parts segmentation.(4)Based on the part segmentation model trained by the improved Deeplabv3+network,this paper uses deep learning framework Pycharm and Tkinter to design a part segmentation system based on the improved Deeplabv3+network,which can achieve the segmentation task of the input part image.