Research On Underwater Croaker Inteilligent Monitoring System Based On Deep Learning

Fishery is a significant industry of the national economy.At the same time,the traditional aquaculture models and methods can no longer satisfy the requirements of fishery intelligence and digital development.More specifically,the underwater behavior characteristics of fish are important references for intelligent breeding.However,there is still no effective method to digitize the behavior of underwater fish.It is possible to realize the intelligent monitoring and digital analysis with the help of deep learning technology in recent years.Therefore,this thesis proposes an improved detection and tracking algorithm which is employed to research underwater yellow croaker.In addition,an underwater yellow croaker monitoring system is created to realize the real-time detection and tracking.The main work and innovations of this thesis are listed as follows:1.In this thesis,a series of improvements are made based on Yolo v5 model to solve the problems in underwater yellow croaker object detection tasks:(1)The swimming shape of fish is irregular in the process of underwater fish detection.In addition,high-density underwater fish will always overlap.However,they will lead to missed and false detection.In order to solve this problem,A parallel cascaded attention mechanism backbone network is proposed.With the help of network reconstruction ideas such as residual connection,low-dimensional embedding,parallel cascade,and attention mechanism,this thesis redesigns the C3 module in the backbone network for purpose of enhancing the high fidelity of the feature which is extracted by the backbone network.(2)Aiming at the issues that the shooting area is limited and the fish target at the image boundary is lost,an adaptive mechanism of bidirectional feature multichannel fusion is proposed.In the feature fusion stage of the neck network,the number of branches of the shallow representation information is increased,and the weight parameters are assigned,then combined with the bidirectional cross-scale connection structure to accelerate the adaptive learning of the network.(3)In order to reduce the feature conflict of classification regression,a lightweight multi-task decoupling output head is proposed.It reconstructs the original network output method and decouples the regression and classification tasks in a lightweight manner in order to reduces the mutual exclusion of prediction information.The average precision of the improved detection algorithm in this thesis can reach 87.1% in the underwater yellow croaker dataset,wich is 5.1% higher than the original algorithm.It reflects the superior performance of the improved algorithm in this thesis.At the same time,the improved algorithm of this thesis is also tested on the URPC public dataset,and the mean average precision can reach 86%,which is 2.0% higher than the original model.It verifies the high generalization of the improved algorithm of this thesis in other aquatic detection tasks in complex scenarios.2.An improved yellow croaker tracking algorithm is proposed.In this thesis,the adaptive Kalman filter algorithm is used to eliminate the blurring of underwater video caused by camera shake.Based on the image registration method,the predicted position is adaptively corrected by affine transformation with enhanced correlation coefficient.In this thesis,the motion trajectory estimation of yellow croaker is realized by compensating the motion estimation error.Based on the improved algorithm in this thesis,the multiple object tracking accuracy of yellow croaker dataset can reach 59.9%,which is 1.1% higher than the original algorithm.3.An intelligent monitoring system for underwater yellow croaker is designed.For the actual tasks,the software development process of the system is divided into requirements analysis,functional module division,system deployment,etc.The system adopts the front-end and back-end separation architecture.The front-end developed with Type Script language adopts React framework and designs the visual interface with Ant Design and Tailwind.The back-end developed with Java language and Spring MVC framework.My SQL version 5.7 was used for the database.The algorithm function module was developed in Python language.Docker container technology was used to isolate the operating environment of each functional layer and deploy it on the server.Finally,the system realized the functions including user login,user management,aquaculture tank management,image analysis,video analysis,data query,etc.