Research On Marine Fish Image Detection And Recognition Algorithm Based On Deep Learning

Fish are an important part of the marine ecosystem.Studying the distribution,quantity,ecological characteristics and biodiversity of fish can help protect the marine ecological environment and resources,promote the sustainable development of fisheries,as well as facilitate the progress of marine scientific research and other fields.Initially,this process required manual sampling and diving observation,and with the development of technology,automated technologies based on image acquisition were gradually applied to marine fish detection and identification.However,the following difficulties arise in the process of marine fish detection and identification: it is difficult to extract the features of fish because their textures,colors and other features are complex and the quality of data taken underwater is poor due to the influence of underwater environment such as light,water quality and occlusions;the morphology and color and other features of the same species of fish vary greatly,and the features are similar between some different fish,resulting in fish Some scenarios require real-time fish detection,such as marine fisheries and aquaculture,so the speed and accuracy of model detection need to be balanced to improve the real-time performance of model detection.In view of the above problems,this paper studies the detection and recognition classification algorithms of marine fish based on deep learning methods.The results are as follows :(1)In order to improve the quality and usability of the dataset,the dataset is processed.First,the desired dataset is obtained from public datasets;subsequently,data cleaning,including operations such as de-duplication and annotation,is performed;preprocessing,such as image scaling and normalization,is performed on the original images;in addition,more training data are generated by data enhancement techniques to improve the generalization ability of the model;finally,the dataset is divided into training and validation sets for model training and testing.(2)In order to quickly detect the location of fish and identify and classify them,a marine fish target detection model based on an improved FCOS network is proposed for marine fish detection.First,the model uses the single-stage algorithm FCOS as the basic architecture,and uses the lightweight Mobile Netv2 as the backbone network to improve the detection speed;second,the adaptive spatial feature fusion module is introduced to avoid the inconsistency of scale features and improve the detection accuracy;finally,the center-ness branch is introduced into the regression branch,and the joint crosscomparison loss is introduced as the optimization objective Finally,the center-ness branch is introduced into the regression branch,and the joint cross-ratio loss is introduced as the optimized objective function to improve the model performance.Through comparison experiments with other models on the marine fish dataset Fish4 knowledge,it is demonstrated that the detection speed and accuracy of the proposed model are improved,and the average detection accuracy of the proposed new model is 99.79% and the detection speed is 22.07ms/img,and the recognition accuracy is significantly higher than other models.(3)In order to solve the fine-grained problem of underwater fish with large background noise,large feature differences between the same species of fish,and similar features between different classes of fish resulting in low recognition accuracy,a bilinear marine fish fine-grained recognition model based on Rep VGG is proposed.The model uses two improved Rep VGG networks as feature extraction networks,and uses the pooling residual module to fuse the features of Rep VGG networks to improve the feature extraction ability of the model for fish,while adding an attention mechanism to improve the learning ability of the network;during the training process of the model,a central loss function is also used to expand the inter-class differences and improve the intraclass compactness,which is more favorable for fish fine-grained identification.By selecting 200 classes with fine-grained features from the large scale fish dataset Wild Fish and comparing them with other models,the results show that this model can effectively improve the accuracy of fish fine-grained recognition.