Research On Underwater Fish Counting Method Based On Improved DeepSORT

In intensive aquaculture,reliable estimation of fish quantity is very important for aquaculture.Traditional fish counting mainly depends on manual statistics,which is time-consuming and laborious,and easy to cause stress damage to fish.The underwater detection method based on machine vision will not pollute and damage the water environment,and will not affect the normal behavior of fish.This paper proposes an underwater fish counting method based on improved DeepSORT.For the collected underwater fish video,this paper proposes(1)an underwater fish recognition method for turbid waters,and selects the best underwater image enhancement method under the experimental scene through experiments;(2)Underwater fish counting model based on improved DeepSORT tracking algorithm.The main contents of this paper are as follows:(1)In the field of aquatic ecology,compared with the current artificial statistical analysis methods used to monitor and evaluate quantity,machine vision provides an alternative method with faster speed,lower cost and higher accuracy.The water quality of pelagic mariculture is good and the visibility of seawater is high.At present,most of the data sets of mainstream research are collected in the seawater environment or under the autotrophic fish tank.In order to solve the problem of fish counting in turbid underwater environment in freshwater or offshore areas,this paper proposes a fish detection method for turbid underwater environment,including: 1)underwater image enhancement technology can be divided into traditional image enhancement algorithm and image enhancement algorithm based on neural network.This paper selects several representative enhancement algorithms,The collected data sets are enhanced respectively,and the advantages and disadvantages of each enhancement algorithm are analyzed combined with subjective visual perception and image enhancement evaluation index;2)For the data set processed by each enhancement algorithm,YOLOv4 is used for training.According to the map value,the best underwater image enhancement algorithm under the specific scene of this paper is selected.(2)DeepSORT is a classic algorithm for real-time multi-target tracking based on detection.During tracking,DeepSORT uses Faster R-CNN to detect the target.Although the detection accuracy is high,the prediction speed is insufficient.This paper proposes a method based on improving the tracking and counting of underwater fish based on DeepSORT,including: 1)replacing the detector of DeepSORT algorithm with YOLOv4,Improve the detection speed while ensuring the detection accuracy;2)In cascade matching,complete intersection union ratio(CIOU)is used to replace intersection union ratio(IOU),so as to improve the accuracy of target matching and achieve higher accuracy of underwater fish counting.Through experiments,it is found that(1)among the eight selected image enhancement methods,the DCP image enhancement method performs best in combination with the map trained by subjective visual perception,information entropy,UCIQE and YOLOv4,and when the training weight file is used for DeepSORT underwater tracking of fish targets,the counting result is the closest to the actual result relative to its enhancement method.(2)In the tracking and counting module,the improved DeepSORT is used to track and count the fish in the underwater video.The results show that the tracking and counting effect is good.The average accuracy of counting on the 20 s video is 91.67%,and the detection speed is 20.83 ms per frame.It is suitable for batch processing of underwater fish video.