Fine Grained Image Classification Based On Multi-scale Feature Fusion And Attention Mechanism

With the continuous development of deep convolution neural network(VGG-Net,Res Net,Densenet,etc.),it is found that the effect of processing common coarse-grained image classification has become more and more accurate.In a large number of classification tasks,it has even exceeded the manual classification.Gradually,many scholars and researchers focus on the fine-grained image classification task.However,the direct use of these depth convolution neural networks for fine-grained image classification can not get good results,mainly because these fine-grained images have many similarities,it is difficult to find the features with obvious differences in fine-grained images by using these depth convolution neural networks directly.In order to meet the requirements of fine-grained image classification,this paper proposes a new fine-grained image classification method based on the traditional deep convolution neural network(VGGNet,Res Net)by improving its structure.The main purpose of this method is to obtain the fine-grained image which can more represent the small class of fine-grained features.Through multi-scale fusion and attention mechanism,the depth classification model is constructed,which can achieve good experimental results.In the process of convolution pooling of the image in the deep convolutional network,some small discriminable regional features may be missing.This article uses end-to-end weakly supervised training.In the traditional convolutional network,a multi-scale feature fusion method is introduced,which combines the shallow features and high-level semantic features of the image well.The attention mechanism is applied to the spatial dimension and channel dimension of the feature map,and the spatial attention module and channel attention module of the model are designed.By distributing the weights on the features,the discriminability area in the fine-grained image is more prominent,ensuring Robust performance of the model.In this paper,the models are constructed from the point of view of series and parallel modules,and experiments are carried out on both modes.During the experiment,the three public fine-grained image data sets of CUB-200-2011,FGVC Aircraft and Stanford Cars were trained and tested and evaluated,and end-to-end experiments were carried out on the improved model network model based on Res Net101.The experimental results show that the highest accuracy of the prediction results is 1.2%,0.8%,1.1% higher than the classification accuracy of other models in the same period,and good experimental results have been achieved.