| The research and development of machine vision technology in the field of intelligent manufacturing has a long history,and already has a considerable system and scale of landing projects.However,the performance of traditional machine vision technology is limited,and it cannot achieve ideal results in large-scale image recognition,high-complexity object detection,and image segmentation.The emergence of deep learning has fundamentally solved the performance problems of traditional algorithms,and has broken the performance bottlenecks of traditional algorithms in both image processing and natural language processing.However,the high performance of deep learning depends on a large number of data sets and deeper model structures,which leads to too many model parameters and excessive calculations,and has a high dependence on the performance of hardware devices.In the actual industrial production field,most of the mechanical equipment belongs to embedded devices or mobile devices.Its memory capacity is small and the CPU computing capacity is limited.It is difficult to meet the computing needs of deep learning models.Therefore,how to make deep learning models run smoothly on embedded devices and meet the real-time requirements in industrial manufacturing has become a current research hotspot,which is called "lightweighting" of models.The application of deep learning in industry is currently in its infancy.Starting from the latest results of deep learning,this subject takes mechanical parts identification in the field of industrial manufacturing as the research object,and researches from three aspects: the collection of raw data,the preprocessing and composition of data sets,and the improvement of lightweight models.In the application of deep learning,it achieved an accuracy of 98.81% on the 100-part data set collected independently.The main contents of the article are: 1.Aiming at the common environmental characteristics in the industrial manufacturing field,comprehensively considering the data requirements of deep learning,a multi-angle data set acquisition system was designed and implemented to collect the data sets required for this topic.2.In order to improve the quality of the data set,to solve the problems of too high image resolution and too large blank information area,based on Gaussian blur and canny edge detection algorithm,a part positioning method based on parameter loop detection is proposed to achieve accurate positioning of parts.The image is then cropped and scaled according to the positioning results,and a standardized data format is constructed.The data set enhancement method is used to move the data set image up and down,left and right,clockwise and counterclockwise,and perform horizontal and vertical mirror operations.Increase the amount of data.3.For the problem that the current deep learning network model is too large to run smoothly in machine vision embedded devices and cannot meet the real-time requirements,based on the Densenet model with high original parameter efficiency,combined with deep separable convolution operations and Feature-dropouted concatenation is a lightweight improvement on Densenet,and a FD-Densenet model is proposed.Experimental results confirm that the model can maintain the accuracy equivalent to Densenet,and greatly reduce the number of parameters and calculations.It maintains a model size comparable to the current mainstream lightweight models and achieves higher accuracy. |
