Design And Implementation Of Fish Individual Recognition System Based On Deep Learning

Accompanied by the rapid development of deep learning technology,object detection and biometric recognition technology based on convolutional neural network in deep learning is widely applied to the work of fish recognition.Accurate individual recognition of fish in farming waters can help aquaculture enterprises better understand the growth of farmed fish.On the one hand,the health condition of individual fish can be better grasped,and timely treatment can be given to the sick fish as soon as they become ill.On the other hand,it can make a more reasonable bait feeding plan according to the growth of individual fish,allowing for the accurate feeding of cultured fish while avoiding bait waste and saving cost for the enterprise.However,issues like blurred images of collected fish individuals caused by low light levels,the presence of water plants and other floating objects,significant size variations between fish individuals,or body shading between fish individuals that affect recognition accuracy have not been effectively solved.The two individual fish recognition approaches that are proposed in this research have a progressive relationship,and they increase the overall performance of the individual fish recognition model by increasing the model’s detection algorithm and recognition algorithm,respectively.The research content of this paper is as follows:1.Blurred images are a common issue in real underwater environments because of the low light levels and the presence of floating plants or other items.To overcome this problem this paper proposes a merger of improved YOLOv4 and FaceNet for underwater fish individual recognition.Convolutional Block Attention model and YOLOv4 algorithm are coupled as part of the model improvement process to increase the network’s capacity for learning critical features.This makes it possible to overcome the aforementioned issues in actual underwater situations and further improves the accuracy while boosting the model’s overall performance.2.Based on the work mentioned above,we propose an improved and fused YOLOv4-tiny and ArcFace technique for detecting individual fish in real underwater environments where there are significant changes in fish size and body occlusion between individual fish.The YOLOv4-tiny and ArcFace algorithms’ network structures are enhanced and optimized,which increases the ability of features to spread throughout the network and,as a result,the deep network’s capacity for feature extraction.In addition to improving the identification model’s overall performance,the model can execute individual fish recognition tasks and produce more precise recognition results for the aforementioned issues.3.This paper builds and implements a deep learning-based fish individual recognition system,which is subdivided into fish individual detection and fish individual recognition modules,based on the aforementioned research.Firstly,we construct a fish dataset that meets the input requirements of the model,and then input the dataset into the fish individual detection module,and generate the prediction frame and coordinate value information of the fish individuals to be detected by the fish individual detection model.The next step is to feed the prediction frame and coordinate value information into the individual fish recognition module to reliably recognition fish individuals and report the identification results.