Coral Fish Detection And Recognition Based On Deep Learning

In the field of marine science,the detection and recognition of coral reef fish plays a key role in marine ecological monitoring and the protection of endangered marine species.Studies have shown that:the richness of coral reef fish variety and activity traces directly reflect the health and biodiversity of coral reef ecosystems.At present,with the significant degradation of most coral reef ecosystems in the world,this research topic has received extensive attention.Driven by the growing demand for underwater ecological monitoring and biodiversity monitoring,the use of advanced imaging systems to collect more marine-related multimedia data and conduct effective analysis has become a problem in the marine sciences.At present,many countries and regions around the world have deployed underwater video surveillance systems in typical coral reef waters to collect and monitor data on biodiversity and ecosystem health near coral reefs.At the same time,a large number of underwater video surveillance systems have also produced a large number of coral reef fish activity images.The analysis and organization of these data often requires very time-consuming input by human observers and requires expert-level knowledge.That greatly limits the analysis and utilization of underwater video data.That greatly limits the analysis and utilization of underwater video data.Therefore,there is an urgent need for automated analytical techniques to assist marine biologists in the analysis of underwater images using exponentially increased underwater data.This paper studies the coral reef fish detection and recognition in underwater and has achieved the following research achievements:Firstly,proposed a method based on feature fusion for coral reef fish detection and recognition,called FFDet.FFDet is a full convolutional neural network for coral reef fish detection based on deep learning,which consists of an SSD-based backbone network.Unlike SSD,it combines shallow features with detailed context information with deep features with rich high-level semantic information,and using fused enhancement features for prediction,while features from multiple layers are used for detection and recognition of fish of different scale.The experimental results in the SeaCLEF dataset show that:FFDet is not only superior in performance to SSD,but also superior to the two other popular end-to-end models in terms of detection performance and speed,especially better overall in large fish detection.Then,proposed a method for detecting coral reef fish that combines optical flow information,called FTDet.In underwater video scenes,affected by coral reef fish posture changes,motion blur,uneven illumination,and occlusion,the appearance of coral reef fish has been significantly degraded in some video frame.This apparent deterioration in appearance has made it difficult for some coral reef fish to be detected or identified.FTDet enhances the object appearance information in the current video image by extracting optical flow information from adjacent video images and then merging the current video image with the video image propagated through the optical stream,thereby achieving better detection,effect.Experimental results in the SeaCLEF dataset show that:FTDet can have better detection results.