Object Detection And Tracking Driven Hunger State Recognition Of Channa Argus

The identification and analysis of Channa argus hunger status is of great significance for the intelligent breeding and breeding strategy adjustment of Channa argus.The non rigid body of Channa argus leads to a large number of intra-class occlusion phenomena during the breeding process,making it difficult to detect and track multiple Channa argus at the same time,making it difficult to analyze the hunger status of Channa argus.With the development of deep learning,algorithms such as Object Detection,Object Tracking,and behavior recognition are increasingly developed,but their application in the field of Channa argus farming is not widespread.Therefore,based on deep learning,this article proposes a method for Object Detection and multi Object Tracking of Channa argus,and analyzes the hunger behavior of Channa argus based on Bi-LSTM(Bi-directional Long Short-Term Memory).By analyzing the degree of hunger,it can help adjust Channa argus breeding strategies,accurately feed to save feed costs,and achieve intelligent fishing.This article mainly conducts the following three aspects of research work:(1)A horizontal cross scale fusion algorithm for Channa argus Object Detection based on YOLOv4 is proposed.The non-rigid body of Channa argus causes diversity of occlusion during swimming,resulting in a low detection accuracy for Channa argus under intra class occlusion.To address this issue,a Channa argus Object Detection(COD)dataset was established.Based on this,a horizontal cross scale fusion Res2-YOLOv4 Channa argus object detection method based on YOLOv4 was proposed,aiming to solve the problem of difficulty in detecting Channa argus under intra class occlusion.First,the Cross Scale Res2 Net backbone network is proposed to obtain effective features more accurately in the case of diverse intra class occlusion;Secondly,Io U Soft-NMS is proposed to use the Io U(Intersection over Union)of real and predicted boxes as the screening criteria for prediction boxes.The Io U score representing positioning accuracy is used as the basis for sorting prediction boxes to improve network accuracy.Using the COD dataset to validate the model,the results showed that AP50 reached 94.77%,AP75 reached 92.20%,which is 4.61% higher than the benchmark network YOLOv4,and AP75 increased by 3.73%.The experimental results confirm the effectiveness of the method and achieve high-precision detection of Channa argus under intra class occlusion.(2)A multi object tracking algorithm for Channa argus based on Byte Track has been proposed.Multiple Object Tracking(MOT)aims to achieve object detection and data association between different frames.However,multi object tracking depends on the effect of detection.Due to high-density aquaculture,the frequent swimming of Channa argus will lead to constant changes in posture and generate a variety of intra class occlusion,resulting in a high rate of false detection and missed detection,making tracking unable to produce complete tracks,leading to fragmentation of tracking.To address the above issues and improve the accuracy of Channa argus multi object tracking,a Channa argus multi object tracking(CMOT)dataset was established.Based on this,a Byte Track Fish Channa argus multi object tracking method was proposed to achieve high-speed and accurate Channa argus multi object tracking based on object detection.In the detection part,YOLOX backbone network is improved by introducing cross scale horizontal fusion network and spatial attention mechanism to extract and fuse richer features.In the data association section,the impact of appearance similarity on tracking is no longer considered,and only the data is correlated and matched based on motion position similarity to enhance the tracking accuracy of snakehead in situations of similar appearance and intra class occlusion.The experimental results show that the MOTA of this method can reach 81.4%,which is 1.2% higher than the benchmark paper Byte Track.It can accurately and effectively achieve multi object tracking of Channa argus.(3)A Bi-LSTM based method for identifying and classifying the hunger status of Channa argus is proposed.The aggregation degree and swimming speed of Channa argus are the main hunger characteristics,and changes in position and swimming speed can reflect changes in hunger status.At the same time,the hunger state of the Channa argus has obvious temporal characteristics and is a gradual process of change.The changes in the hunger state of the Channa argus within a single frame image and a single video sequence sample are not significant.Bi LSTM bidirectional short-term memory network has memory cells and gating mechanism,which can selectively forget and learn information and effectively realize sequence feature learning.Based on the results of multi object tracking,extract the position features and swimming speed of Channa argus,construct a sequence of Channa argus hunger features,establish a dataset of Channa argus hunger level(CHL),and propose a Bi-LSTM-Fish algorithm for identifying Channa argus hunger status to achieve classification of Channa argus hunger status.The experimental results show that this method can effectively identify and classify the hunger status of Channa argus,which is beneficial for precise feeding and cost saving of Channa argus aquaculture feed.