Freshwater Fish Image Classification Based On Convolutional Neural Network

The identification of freshwater fish species is a key step in the planning of fishery resources development and the development of fisheries to automation and intelligence.It is widely used in fishery resources management,freshwater fish aquaculture,freshwater fish sorting and deep processing.The research of freshwater fish species identification combined with deep learning technology is of high value.Therefore,aiming at the problems of slow speed,low accuracy and difficulty in feature extraction of traditional freshwater fish classification methods,this paper carries out the research of freshwater fish image classification based on convolutional neural network.The main research contents are as follows:(1)A dataset of freshwater fish images required for model training was constructed.Firstly,the images were collected by handheld device field shooting and Internet image collection,and the freshwater fish images were sorted and classified.Ten kinds of freshwater fish images including grass carp,bighead carp and carp were collected and sorted.Secondly,from the perspective of the insufficient number of freshwater fish image samples,a series of data enhancement operations including vertical flipping,horizontal rotation of a certain Angle,and random cropping are performed on the training data set,which improves the overfitting situation of model training to a certain extent.(2)Aiming at the complexity of freshwater fish images and the difficulty of feature extraction,this paper focuses on an improved Xception model.Firstly,hierarchical class connections are constructed in the depthwise separable convolution module of Xception model,and the feature maps are grouped and filtered by channel to extract multi-scale features in a more finegrained way and increase the richness of feature information extracted by each network layer.At the same time,1 × 1 standard convolutional layers are added before the depthwise separable convolution module for channel grouping to further strengthen the information interaction between different channels of the feature map.The influence of different number of groups on the classification performance of the model is studied through experiments,and the best number of groups is determined.At the same time,the influence of different hyperparameters on the classification performance is discussed.Experimental results show that the proposed method can identify freshwater fish species with high accuracy.(3)Aiming at the problem that freshwater fish images are highly similar and difficult to identify,this paper proposes a classification method for freshwater fish based on Res2Net50.Firstly,the SA module in SANet is embedded into the residual module of Res2 Net,and the spatial and channel attention feature information of different groups is fused to realize the recaliation of feature information.Secondly,an average pooling layer is added to the downsampled residual connections,and three 3 × 3 convolution kernels are used to replace the7×7 convolution kernel of the first convolution layer in the Res2 Net model,which not only strengthens the nonlinear ability of the model but also reduces the loss of feature information in the downsampling process.Finally,CELU activation function was selected to improve the expression ability of the model.Experimental results show that the proposed method can effectively classify freshwater fish.(4)In order to solve the problems of slow speed and large amount of calculation in the process of freshwater fish species identification,this paper proposes an improved GhostNet model.The SE module in the GhostNet model is improved,and the influence of compression rate and feature fusion position in the module on the classification performance of the model is explored,so that the model can better learn the importance of different channels of the feature map.Compared with other lightweight models,it is verified that the proposed method has better classification performance while maintaining the computational complexity basically unchanged.