Research On Fine-grained Image Classification Algorithm Based On Attention Mechanism And Feature Interaction

With the development of artificial intelligence technology,traditional image classification cannot meet the needs of actual scenes.Fine-grained image classification technology has become an important tool in daily life and industrial production.Aiming at the problems of large intra class difference,small inter class difference,complex background and pose difference of fine-grained images,this paper mainly uses two strategies of feature interaction and attention mechanism to optimize the classification network model and improves the accuracy of network classification.The specific research is as follows:(1)Aiming at the shortcomings of B-CNN algorithm which lacks spatial capture and channel correlation,an improved B-CNN algorithm based on attention mechanism and feature interaction is proposed.The algorithm retains the advantage of bilinear pooling to improve the ability of feature expression.Before bilinear pooling,the attention module and feature interaction module are added to construct spatial information and channel interaction,so as to highlight the distinctive local areas in the expression image.(2)Aiming at the problem that the current fine-grained image training data is insufficient and the traditional data enhancement technology is not suitable for fine-grained images,a fine-grained image classification algorithm based on complementary data enhancement interaction is proposed.Firstly,a complementary attention data augment module is constructed based on the attention map to obtain the erased image and local cropped image,and then the feature interaction module is used to capture the comparative features between the original image and the augmented image,learn the complementary information between the images,and obtain a richer feature representation.(3)Aiming at the problem that it is difficult to locate the differentiated parts of the target in fine-grained image task,a fine-grained image classification algorithm based on multi-stage feature enhancement interaction is proposed.In order to generate differentiated component features,the feature enhancement suppression module is constructed to enhance the most prominent part of the feature mapping in the current stage,so as to obtain the representation of specific components and suppress it,forcing the next stage to mine other potential features,and then the part feature interaction module is constructed by using enhanced part features to enrich the feature representation of each component by aggregating other component information.To sum up,this paper mainly uses the improvement measures such as attention mechanism,feature interaction and data enhancement to optimize the network model,mine the feature representation suitable for fine-grained classification tasks,and improve the accuracy of fine-grained image classification.A large number of experiments show that the performance of the proposed method is significantly improved and has practical application value.