Research On Object Recognition Method Based On Deep Learning And Its Application In Industrial Scene

Object recognition is the basis of high-level computer vision applications.Its main task is to locate the location of a specific object and determine its category information from the video or image data obtained from the visual sensor,usually in the form of a bounding box to identify the object in the image.specific location and output category information and its confidence above it.As one of the important technologies of robot visual perception,object recognition enables the robot to realize a series of functions such as image recognition,text recognition,and object detection.With the development of industrial strength and technology,the performance and functional requirements of industrial robots are also getting higher and higher.Therefore,studying the performance of various aspects of object recognition methods has theoretical guidance and reference significance for improving and improving the perception ability of robots in complex environments.In order to make industrial robots adapt to work in complex environments,optimize and improve the speed and accuracy of object recognition technology is a necessary and practical work.In this paper,combined with the challenges of object recognition in industrial scenarios,the research on deep learning object recognition technology is carried out from three aspects:real-time,recognition accuracy and robustness.Firstly,the theoretical method is researched on the network public data set under the background of urban traffic similar to the industrial scene,and then the theoretical research is verified in the self-built industrial scene dataset,and the optimization of the object recognition technology in the industrial scene is realized.The main research contents of the paper are summarized as follows:First,in terms of real-time performance,the lightweight optimization method of the YOLO v3 model in the urban context is studied,and experiments are carried out on the "Urban object detection" dataset through channel pruning,layer pruning and hybrid pruning of the two.It shows that the two separate pruning strategies of channel pruning and layer pruning can reduce the weight of the model under the premise of ensuring accuracy,and the hybrid pruning of the two can make the pruning degree of model parameters and model volume less accurate.conditions can be further improved.Then,for object recognition in continuous scenes,an object recognition method that fuses time series information and scene prior information is proposed,which utilizes the time series correlation between consecutive object frames and the prior knowledge accumulated in human daily life,which realizes the decision-level fusion of time series information and scene prior information.Experimental studies show that this method can better suppress the influence of interference factors such as illumination conditions and object scale changes,and significantly improve the comprehensive performance of object recognition.Finally,the application of the above two methods in industrial scenarios is realized.The existing network public datasets about industrial scenarios are few,which is not conducive to the development of research work.The author collects industrial scenario data sets produced by self-made machines,and selects three types of basic object in the scenario for technical verification and analysis.Experiments show that the model lightweight optimization and information fusion methods also have good applicability in industrial scenarios.