Research On Tuna Identification Method Based On Deep Learning

All the data of this research comes from the research and development project of tuna longline fishing of Sinohydro Group Oceanic Co.The use of computer vision to identify tuna is as a very important research topic in the intelligent process of ocean fishing tuna,which can provide a theoretical basis for the status of tuna resources in different sea areas and provide relevant technical support for further realization of intelligent fisheries.In recent years,with the update and iteration of deep learning algorithm,this paper proposes a deep learning based tuna identification algorithm model tuna-YOLO,adopts the target detection algorithm based on lightweight network to achieve model acceleration,and uses the local feature information of tuna and model compression technology to improve the network model performance,and relies on the target tracking algorithm to achieve the number of tuna statistics.The main work of this paper are as follows:(1)The video taken by electronic surveillance is extracted to the dataset containing tuna in separate frames,and image pre-processing is used to improve the image clarity and obtain more tuna data for the case that the dataset is blurred at night and the amount of data is sparse,and the image after image pre-processing can effectively improve the accuracy and generalization of the target recognition algorithm.(2)The traditional target detection network and deep learning target detection network are compared and analyzed,and the YOLOv3 target detection algorithm is established as the base network model by combining the application scenarios,advantages and disadvantages of target detection algorithms and algorithm evaluation indexes,and the performance of different lightweight networks as the feature extraction network of YOLOv3 is explored,and the performance of different lightweight networks is analyzed to determine the Mobile Netv3 as a feature extraction network module to improve the real-time detection efficiency of the algorithm.(3)Due to the blurred background and unclear texture features of tuna,the original attention mechanism SENet of Mobile Netv3 was also replaced with CBAM and ECANet in order to enhance the extraction capability of tuna feature information,and CBAM was selected as the attention mechanism module of Mobile Netv3 for the comparative analysis of different attention mechanisms.In comparison with the unimproved YOLOv3,the number of its middle parameters is reduced from 234.74 MB to 88.45 MB,the number of floating point operations is reduced from 32.767 G to 8.676 G,and the overall frame rate(fps)of the algorithm model is improved to the extreme.The improved network model tuna-YOLO(Mobile Netv3 with CBAM and YOLOv3)was passed through the teacher network model Dense Net201-YOLOv3 for knowledge distillation,and the accuracy of the tuna recognition network model was improved,and the m AP reached85.74%.(4)In order to realize the tuna catch statistics,the improved tuna identification algorithm model is used to obtain the regression frame information of each frame in the video,and the target tracking algorithm based on IOU to track the tuna will have problems such as artificial occlusion that leads to the loss of the target tracking object,and the Kalman filter-based target tracking algorithm is proposed to achieve the statistics of tuna catch.(5)Using Py Qt5 framework to design the visualization interface of tuna identification and catch statistics,two functional buttons of target identification and target tracking count on the interface are used to interact with the corresponding algorithm,target identification shows the category and location information of tuna,and target tracking count interface shows the quantity information of different categories.