Research On Fine-grained Image Classification Based On Two-stream Network Model

The problem of fine-grained image classification has always been a challenging research topic in the field of computer vision.Conventional image classification methods are often classified objects with distinct and discriminative features.For example,the classification of birds and dogs,the classification of cars and airplanes,etc.,the distinction of such objects often requires only some shallow features to complete the classification task.Traditional coarse-grained image classification requires a large amount of manually labeled image data,and manual labeling of each image undoubtedly consumes a lot of manpower and time.With the development of society and the advancement of science and technology,people are not satisfied with the classification of coarsegrained images,and the classification of fine-grained images is more and more relevant to people's needs.Objects classified by fine-grained images are subclasses of a certain class,for example,classification of different species of birds,classification of different species of dogs,.Due to the subtle differences between different subclasses,there are large differences in subclasses.In addition to the object itself,the interference of illumination,shooting angle,and background information affects the classification effect to some extent.Classification of granular images is more difficult.Fine-grained classification has greatly improved the efficiency and accuracy of classification from the traditional methods of relying on artificial methods to the methods of deep learning.In the method of fine-grained image classification using neural networks,the classification network model is also developed from a shallow network model to a deep network model,and the width is also developed from a single network model to a dual-stream network model or even a threestream network model.The dual-flow network model has obvious advantages in classification depth and width compared to traditional methods in classification accuracy and efficiency.In this paper,based on the method improvement of the dual-stream network model,in view of the great success of the ResNet network model in the field of image classification,firstly replace one of the single-flow networks with the residual network based on the original dual-stream network,and form a residual hybrid network model to further enrich The feature extraction is performed,and then a four-stream network model is formed by superimposing a dual-flow network model on the basis of the residual hybrid network model.The additionally superimposed dual-flow network model is composed of two different depth VGG network models to further Improve and enrich theextracted features.The experimental results verify the effectiveness of the proposed method.