Image Classification And Target Detection Of Reef Organisms In Marine Ranch Based On Deep Convolution Neural Network

Status and change of biological resources in marine ranch can be recorded and reflected through underwater video.Therefore,it is needed to develop the image-based biological classification and target detection methods,so as to realizing the full potential of machine vision technology in the study and regulation of biological communities in marine ranches.In this study,we collected underwater videos with reef,algae bed and sediment background in Yantai and Weihai area of North China,and compared the effects of various color compensation methods,such as,color compensation based on green channel,contrast limited adaptive histgram equalization,and so on.Classification image dataset for reef biological is established by manual annotation,there are 11 species of common fish in reef area,such as Lateolabrax japonicus,Liza haematocheilus,Sebastes schlegelii etc,3 species of echinoderm and1 species of crab,with a total of 23211 images.Using Paddle Paddle2.0 and Paddle X development kit,transfer learning image classification networks Alex Net,Mobile Net V3 and Res Net50,verify the robustness of the algorithm on underwater images with noise,the accuracy rate of applying category to test set is 96.64%,94.75% and 99.23%respectively.Resnet50 model has better robustness in the image dataset with Gaussian noise.The underwater video of Muping Yunxi sea area from August 2019 to April 2020 is selected.1856 images labeled by labelme and the corresponding JSON file are transformed into rectangular ones by extracting contour corners.The target detection data set containing 12 kinds of reef fish and 6405 labels is established.The experimental system was windows10 operating system,python3.8 programming software,and Baidu Paddle Paddle 2.0 Paddlex development kit and cuda11.0 are used as the development platform.Transfer learn yolov3 model based on darknet-53 feature extraction architecture.The model is trained,verified and tested through parameter adjustment,and the best＿model predicted64.48% map for the target detection data set.The results show that the computer vision technology based on deep learning has great application potential in the video monitoring technology of marine ranch,and can provide new ideas and methods for the monitoring and management of marine ranch biology in China.