Fine-gained Fish Image Classification Based On Adversarial Learning

Deep-sea fish are a vital part of the Marine biological systems,it is indispensable to human basic material guarantee,at the same time play a key role in the maintenance of ecosystem balance.Detection of Marine fish resource development in the process of the recognition of fish is indispensable.But deep-sea conditions is complex,all kinds of fish forms usually have high similarity between different species,same species of fish also has certain differences,these fish for accurate identification will be a big challenge,is likely to produce misjudgment caused serious economic loss.Therefore,the effective identification of marine fish species will have a great impact on the development of marine resources in China.With the rapid development of artificial intelligence technology,deep learning technology is widely used in various fields,and has achieved quite good achievements.Therefore,on the basis of deep learning technology,this paper carries out relevant research on image classification of fine-grained fish based on adversarial learning.Specific work contents and achievements are as follows:1.Data augmentation method based on generative adversarial network.This method generates pseudo-fish sample data by generative adversarial network,Because there are more species in the fine-grained fish data set,but the sample size of each type is small,which is a typical small-sample classification problem,and the data needs to be expanded.So using the generated attention against network to generate the fish samples.Attention mechanism driven by long-range dependence to model,can obtain the fish details of all the elements of points to generate high resolution characteristics,so as to improve the quality of the generated sample.Compared to the traditional data enhancement approaches such as rotation,scaling,translation,the method can more rich increase the diversity of samples,thus improve the generalization of the model.2.Fine-grained fish image classification based on adversarial bilinear network.This method is expanded on the basis of 1.Due to the high similarity between classes and large intra-class differences of fine-grained fish images,it is difficult to obtain a good classification effect by using the more common image classification method.So using bilinear network as the basic network framework and improving it.Through the characteristics of different layer separately cross product integration,can obtain more abundant and accurate characteristics.At the same time,the loss function is improved to reduce the tendency of overfitting and further improve the performance of the model.3.Fine-grained image classification base on adversarial bilinear residual attention network.In order to further improve the recognition accuracy of fine-grained fish images,this method is further improved on the network structure of 2.The model of characteristic function replacement for feature extraction ability is stronger,the depth of the residual network and in its tail add channels at the same time attention and spatial attention mechanism.It enhances the weight of discriminant local features in fine-grained images and further improves the recognition ability of the model.