Research On Industrial Part Recognition Method Based On Deep Learning

Parts are the basis of manufacturing machines,and the research of industrial parts identification can promote the intelligent development of manufacturing industry.The traditional part recognition method is vulnerable to the influence of light and noise,and is not suitable for large-scale and highvolume recognition tasks.The deep learning target recognition method can avoid the shortcomings of the traditional method,break through its performance bottleneck,and improve the speed and accuracy of part recognition.However,due to the deep network structure of the deep learning target recognition method,it is not suitable for popularization and application due to its large amount of parameters and large amount of computation.Aiming at the problems of low accuracy and slow speed in the recognition of industrial parts by traditional methods,this paper designs an improved deep learning target recognition method.At the same time,combined with multi-target tracking technology,it realizes the intelligent recognition classification and counting of industrial parts,which provides technical support for the intelligent development of manufacturing industry.The main research contents of this article are as follows:(1)In view of the lack of data sets on industrial parts in the existing public data sets,six common industrial parts are selected to construct data sets.Simulate various situations that parts may encounter in the process of recognition,use the camera to take 1400 photos of industrial parts,expand the number of 4600 photos through the methods of brightness,exposure and rotation data enhancement,and use the Labelimg visual image annotation tool to produce the industrial parts data set.(2)Aiming at the accuracy,real-time,and applicability requirements of industrial part recognition,an improved YoLoV4 algorithm for identifying industrial parts was proposed.Use K-means++clustering algorithm instead of Kmeans clustering algorithm to optimize the selection strategy of clustering centers,avoid the result falling into local optimization,and generate new anchor boxes that match the identified parts.Applying the Ghost Net network as a new backbone network to the YoLoV4 algorithm significantly reduces the amount of algorithm parameters to reduce the computational power required for algorithm operation,improves the ability of the algorithm to extract feature information,and further reduces the amount of algorithm parameters by replacing ordinary 3×3 convolutions with deep separable convolutions.While ensuring the recognition accuracy of industrial parts,it increases the recognition speed and the application range on embedded devices.(3)Based on the Deep-Sort multi-target tracking algorithm,an improved Deep-Sort algorithm is proposed for the counting of industrial parts.The DeepSort algorithm is combined with the improved YoLoV4 target recognition algorithm,and the industrial part data set is used for training,which improves the real-time performance of the tracking target,and realizes the tracking and counting function of multi-class parts.