Industrial Parts Recognition Based On Deep Learning In Intelligent Manufacturing

In the process of intelligent manufacturing,the recognition of product parts is gradually transforming from the traditional low-efficiency manual detection to the high-efficiency intelligent contactless machine vision detection.The machine vision detection performance of existing industrial products is also generally affected by the factory environment light and noise,and the robustness and versatility are low,which seriously restricts the high-efficiency intelligent contactless machine vision detection in electronics,automobile,medicine and other fields of application and promotion.Through the analysis of the application status of deep learning technology in product part detection scheme,the problems of low accuracy and low computational efficiency are summarized in this paper.Therefore,an optimization algorithm based on YOLO v3 is proposed to realize the efficient and accurate recognition of industrial parts.The main research contents are as follows:1.At present,there is no special data set for industrial parts inspection.This paper establishes the first data set for industrial parts.The image in the data set is all made by members of our research group.Three-dimensional graphics data set,switch data set and high-speed rail parts data set constitute the industrial parts data set.The data image acquisition methods are camera shooting and web crawler.We also do data processing,annotation,enhancement and other operations on the image.2.Aiming at the problem that the accuracy of YOLO v3 algorithm for industrial parts recognition is low,this paper continues to put forward the following improvement measures.(1)Replace Leaky ReLU with Mish activation function,it can express deeper information and spread information better.(2)Adding SPP pooling layer,SPP connection can increase the extraction of local and global features,increase the receptive field,and enrich the expression ability of feature map.(3)Use k-means++ clustering algorithm based on IOU value to re cluster the data set.(4)Replace BN normalization with CBN normalization,and the CBN shows that more stable characteristic of BN algorithm in small Batch Size.Experiments are carried out on VOC2007 data set.The results show that the recognition accuracy of the optimized method is significantly improved,and the algorithm has good stability.3.Aiming at the low efficiency of YOLO v3 algorithm for industrial recognition,this paper proposes the following improvement measures.(1)Using Densenet+residual model to replace ResNet to avoid over fitting of small data sets.(2)In multi-scale feature extraction,only two sets of multi-scale features are extracted to predict parts,that is to extract six pre-selected frame parameters for training,which can effectively reduce the calculation of data.(3)The location loss calculation is improved based on CIOU loss to reduce the error in calculating the loss function.Finally,the SPP-Dense-YOLO network model is proposed.The model is tested on the classic data set VOC2007 and the self-built industrial parts data set.The experimental results of the model are better than those of the same kind in the comparison of multiple performance indexes such as mAP,Precision,Recall and Fps.It can be effectively applied to the intelligent manufacturing process to improve the recognition effect of industrial parts,improve product quality.