Recognition And Application Of Marine Fish Images Based On Deep Learning

Nearly two-thirds of the earth where human beings live is covered by the ocean.There are abundant natural resources and huge biological resources in the vast ocean.Fish,as the most important kind of marine living resources,is not only an important source of food for human beings and an important material basis for the sustainable development of human society,but also a vital part of maintaining the ecological equilibrium of the earth.All countries in the world attach great importance to the development and utilization of marine fish resources because of its ornamental value,edible value and high medicinal research value.China has abundant marine fish resources,with more than 3000 species which occupy around 20% of the world's fish species.It is even more necessary to protect and exploit fish resources effectively.In the process of exploiting and exploring marine fish resources,it is necessary to identify all kinds of fish.However,different fish species have different shapes,sizes,and textures.At the meantime,different species of the same kind of fish usually have similar shapes,sizes,and textures,which may lead to serious economic losses.Therefore,the study of marine fish identification technology has important academic and economic value for the development and utilization of marine fish resources in China.With the rapid development of computer information technology,deep learning has made important breakthroughs in many fields such as computer vision.Traditional machine learning methods are gradually being replaced by methods based on deep learning.In this paper,deep learning is utilized in the field of marine fish identification,and the research of marine fish identification based on deep learning is carried out.The specific contents include:(1)Fish object detection method based on improved Faster RCNN.Aiming at the problems of the difficulty and low accuracy of existing fish detection algorithms for different sizes and small objects,a fish object detection method based on improved Faster RCNN is proposed.Based on the original Faster RCNN,the new method increase the number of anchors and the multi-channel full connection and deformable convolution which are used to ensure the feature-extraction at different scales and shape,thereby enhancing the robustness of the network to detect objects of different sizes.(2)Fish object segmentation method based on separable residual convolution neural network.Aiming at the problems of low segmentation accuracy,poor effect and low real-time performance of existing fish object segmentation methods,a fish object segmentation network model with better accuracy and real-time performance is proposed,which combines the advantages of residual learning and separable convolution structure in convolution neural network.On the basis of Full Convolutional Neural Network,firstly,the separable convolution structure is used to replace the conventional convolution structure of the original model,which greatly reduces the parameters of the model to be calculated,and then improves the computational efficiency of the model,The dilated convolution is used to increase the receptive field and improve the inference speed without increasing the complexity of the network.Secondly,residual learning is introduced into the model,which combines the deep abstract features with the shallow surface features to learn,so as to locate the gray level category information of the image.While ensuring segmentation efficiency,the segmentation accuracy of the model has also been greatly improved.(3)Fish identification method based on transfer learning and joint loss.In order to solve the problem of low accuracy of training depth model caused by too few fish labeling data sets and poor discriminatory ability of feature output caused by using only Softmax loss function for feature classification,a fish recognition method based on migration learning and joint loss is proposed.Firstly,the fish data set in ImageNet is used to pre-train the deep network model,and the object data set is used to fine-tune the network parameters,then the fish identification model with initialization parameters is constructed.Then,the model is trained and optimized based on the joint loss composed of Softmax loss and center loss,so as to improve the discriminant ability of the output characteristics of the model.(4)Based on the research and work above,a mobile marine fish identification system based on Wechat applet is designed and implemented,and the system is tested through the actual scene.The test results show that the system can detect and classify marine fish accurately,run fast and be easy to use.