Research And Application Of Fish Detection And Counting Based On Deep Learning

Automated intelligent fish farming is the future development direction of aquaculture industry.Fish detection and quantity statistics are an important part of automated aquaculture system.Image processing technology can directly detect and count fish in monitoring video,provide decision-making data for fishery management,and improve the management efficiency of aquaculture industry.When testing fish and estimating biomass in aquaculture,real-time detection and counting of fish stocks can better help farmers accurately calculate reproductive rates and estimate production potential,and guide survival assessment,culture density control,and fish marketing management.Traditional aquaculture farms mostly use manual counting methods.Due to the large number of fish breeding types and high density,it is difficult for farmers to classify and count fish,which requires huge manpower and material resources.The fish detection and counting model can avoid the negative effects of low efficiency and time-consuming,error-prone and easy damage of fish body by manual counting.In addition to aquaculture,fish classification and statistics are also needed in natural waters.China’s oceans and rivers are rich in biological resources and are the largest supply bases for fish,so fish detection and counting are extremely important for aquaculture and fishery resource management.It can not only reduce the mission cost,but also provide data support for scientific research.Therefore,there is an urgent need for a technical implementation that can automatically,accurately and efficiently achieve fish school detection and statistics.As one of the one-stage object detection algorithms,YOLO algorithm combines object location prediction and object category prediction into a neural network model to achieve high accuracy and efficiency of object detection,which is more suitable for practical application scenarios.In order to improve the accuracy of fish detection and counting,this paper studies the YOLOv5 algorithm,which is prominent in the YOLO series of algorithms.By adding an attention mechanism,it can improve the detection accuracy in complex scenes.In this paper,the following corresponding researches on fish detection,tracking and statistics are made:(1)In terms of fish school detection,this paper proposes an improved YOLOv5 fish school detection algorithm.A self-made fish data set was made,and2908 photos collected were manually labeled,with a total of about 15,000 fish.It was divided into a training set and a test set for later training model use.Aiming at the problem of poor detection effect of small target fish school,by adding two attention mechanisms respectively,it is found that after adding SENet module,the comprehensive index mAP@.5:.95 is improved by 3% compared with the original model.After adding the CBAM module,the comprehensive index mAP@.5:.95 is7.3% higher than the original model.The transmission frame per second(FPS)can reach 30 frames per second,which can meet the requirements of real-time detection.(2)In the aspect of fish tracking,the fish tracking model is established based on Kalman filter algorithm.The IOU distance is used as the evaluation matrix,and the Hungarian algorithm is used to process the matching relationship between the fish detection frame and the tracking prediction frame.The detection model combined with tracking algorithm effectively suppresses the missed detection phenomenon of fish school.(3)In the aspect of fish statistics,the fish detection model fused with tracking algorithm is used to study the counting accuracy of small fish fry in the video.Three different fish densities were set,20,50 and 100 small fry were put into the same container for detection,and the accuracy of counting was counted and analyzed.When the density was 20 fry,the counting accuracy of the original YOLOv5 model,the added SENet model and the added CBAM model were all100%.When the density was 50 fry,the counting accuracy rates of the three models were 96%,98% and 98%,respectively.When the density is 100 fry,the counting accuracy of the three models is 91%,95% and 96%,respectively.