Research On Recognition And Counting Method Of Shrimp Shelling Based On Deep Learning

According to the "China Fishery Yearbook" data,as of the end of 2017,Chinese total aquatic product output reached 6,445 tons,the actual processing output was 2,196 tons,and the processing loss rate reached 18.05%.The actual processing output of shrimp is 52.01 tons,accounting for only 2.37%of the actual processing output.It shows that Chinese emphasis on the shrimp processing industry is lower than that of other aquatic products,so the shrimp processing industry has more development potential.Although some prawn processing plants have invested in mechanical sheller equipment,they still rely on traditional manual shelling in actual production and have idle large-scale shelling units.The reason for this phenomenon is that the shrimp sheller equipment in China is still in the early stage of development,and there is no unified evaluation standard for its performance.After comprehensive comparison of manual shelling,the shrimp sheller has no obvious advantages in all aspects.It can be seen that the shelling efficiency and shelling quality of the unit havenot reached the design expectations.It is necessary to improve the shrimp shelling machine to improve the shelling performance and reduce the operating cost.Mastering the performance of the current prawn shelling machine is the prerequisite for improvement,and the statistics of the number of fully shelled and incompletely shelled shrimp are the basis of evaluation.Therefore,this paper uses statistics on the fully shelled and incompletely shelled shrimp produced by the shelling machine.The number of prawns is used as one of the indicators for evaluating the performance of the sheller.Due to the fast-running shrimp processing production line,humans cannot accurately count the number of large quantities of fully and incompletely peeled shrimp.Traditional target detection algorithms cannot achieve fast real-time detection.The target detection algorithm based on deep learning provides a solution to this problem.Program.Through the research of the target detection algorithm based on deep learning,this paper concludes that the detection model based on the YOLO V4 algorithm has the best performance in this detection task.The specific research content and results are mainly as follows:(1)Constructing an ideal data set is a key step in a deep learning-based target detection algorithm.Due to the fast-running production line,individual prawn morphology,shrimp distribution and other factors,it is very difficult to directly obtain the ideal data set.Therefore,this paper screened out 94 valid data images,and carried out image enhancement based on environment simulation,including data enhancement,environment simulation,and noise simulation.A total of 3345 prawn images were generated and a data set was constructed.(2)After theoretical analysis,there are two types of target detection algorithms based on deep learning,two-stage and one-stage.Although the two-stage method has higher detection accuracy,its real-time performance is compared with the one-stage method.It has no advantages and cannot be applied to real-time scenarios.Therefore,this article chooses the target detection algorithm based on the deep tearning one-stage method,which specifically includes SS.D inception V2,SSD mobilenet V2,YOLO V4,YOLO V4 tiny,YOLO V4 tiny-31,CsResNet-PANet-SPP,YOLO V3,YOLO V3-SPP,YOLO V3 tiny,YOLO V3 tiny-prn,YOLO V3 GIOU-12,EfficientNet_b0,YOLO V3 openimages,there are 13 models in two categories.By comparing the performance of the two models of SSD with YOLO V3 and YOLO V4 in this task,it is concluded that the YOLO algorithm is more suitable than SSD.By comparing models other than the SSD model,it is concluded that YOLO V4 has the best detection effect in this task,with a detection accuracy of 99.10%.(3)This paper has developed a shrimp processing detection system,which allows users to choose between online monitoring and offline detection of shrimp peeling status.During online monitoring,users can not only detect prawns in real time,but also establish a connection with the production line to realize the operation parameter monitoring and parameter error real-time statistics of prawn sheller equipment.When offline detection,users can detect a single image or video segment.