Research On Fine-grained Image Classification Algorithm Based On Attention Mechanism

In recent years,with the rapid development of artificial intelligence technology,there has been a breakthrough in the field of image recognition.Fine-grained image classification task is one of the hot research directions in the field of image recognition.Fine-grained image classification refers to the classification of a large category of image libraries or sub-categories in image sets,such as the classification of different kinds of flowers.How to accurately locate objects and extract more expressive features is one of the basic problems of fine-grained image classification algorithms.Fine-grained images often contain a large amount of background information,and the differences between images usually exist in a small area of components.However,classification using only the features of components may ignore the overall features of objects and the relations between components.To solve the above problems,this paper combined the attention mechanism and B-CNN model,proposed a method named as Based on Attention Mechanism in Bilinear CNN(BAM B-CNN).The main research contents of this paper are as follows:(1)An image classification algorithm based on attention mechanism is proposed.The secondary attention model is improved by combining SCDA algorithm and RPN module,so that the object images and component images contain less background information,and the influence of background information is reduced.The channel attention model is used to improve the B-CNN model and extract features,so that the model can learn the nonlinear relationship between channels,and the extracted features are more expressive.(2)BAM B-CNN has three classification subnetworks,which extract the features of input image,object image and component image respectively.Through the classification module,the three groups of features are effectively combined.The effectiveness of the classification module is verified by experiments,and the classification accuracy is improved.(3)Considering the classification algorithm is composed of three classification subnet,the classification accuracy with lower parts classification of network was improved,combined with MA-CNN algorithm to improve component image positioning module,using the improved B-CNN model of parts image training,through improve the classification performance of parts classification subnet,improve the overall classification performance of the model.Experiments on CUB-200-2011,FGVC-Aircraft and Cars are carried out to verify the effectiveness of the proposed algorithm.