Research On Target Detection Of Machine Parts Based On Deep Learning

As one of the key technologies of intelligent manufacturing,part target detection and recognition has aroused widely attention,the advantages of deep learning algorithm in feature extraction provide a novel approach for target detection of mechanical parts.The thesis main research is focusing on the target detection of mechanical parts.For the problems of the traditional mechanical part feature extraction algorithm,such as large detection error and low recognition accuracy,relying on Faster R-CNN network and YOLOv3 network,which are two typical representative networks of deep learning target detection network based on candidate regions and regression,two kinds of object detection algorithms based on deep learning is proposed to improve the size difference,occlusion,low accuracy in part detection,and the shortcomings of manual detection.The main work can be listed as follows:(1)The basic knowledge and three classical deep learning models are introduced,respectively,the related techniques and typical target detection algorithms based on deep learning are summarized with a comparative analysis of their positive and negative.(2)Combining the dimension reduction of feature graphs with the feature extraction and splicing of multi-scale convolution kernels,a multi-scale feature extraction structure is proposed to improve the Faster R-CNN algorithm based on candidate regions,the new net structure reduces the amount of parameters in the network training process effectively and solved the problem of large size difference between different types of mechanical parts.(3)By concatenating the low-level feature map with the high-level's,the local edge details of the low-level feature are integrated into the overall spatial information of the high-level feature,and the utilization rate of feature extraction is improved.Meanwhile,A joint repulsion loss function is proposed to improve the detection accuracy and the low recall rate result from occlusion between dense parts and parts effectively.(4)After multi-scale research on target detection network,it is found that the comprehensive performance of YOLOv3 network structure extraction features with four scales is better than the original one.Using the spatial texture information of low-level features and the semantic information of high-level features together could increase the accuracy of mechanical parts detection.At the same time,the introuduced k-means clustering to cluster and analyze the candidate boxes of the parts could obtained the optimal number of anchors.(5)Considering the influence of illumination,background,occlusion,size,and the other factors on the target detection in actual engineering,a mechanical parts database consisting of five different categories,10000 parts images and their labels to simulate the complex engineering environment is established,which used for the training and testing of the network proposed in this thesis.The results suggest that the improved Faster R-CNN and YOLOv3 target detection algorithms can be successfully applied to the detection of mechanical parts,compared with the traditional detection algorithm,the detection accuracy is higher and the time is faster,which significantly improves the scale of the parts detection,low feature utilization,occlusion and other issues.By comparing analysis of the two improved network performance,the YOLOv3 network is more practical for actual engineering production.