Research On Few-Shot Fine-Grained Image Classification Based On Attention And Local Information

Fine-grained image classification refers to the problem of distinguishing extremely similar subcategories within the same basic category,which has attracted widespread attention in recent years.However,due to the difficulty in collecting data or expensive annotation for some categories in fine-grained images,and the high similarity between classes in fine-grained images,their classification difficulty is higher than that of general image classification.The existing methods have effectively solved the problem of fine-grained image classification,but they require a large amount of data-driven training,so they cannot complete classification well in few-shot scenarios.Based on the above issues,some scholars have proposed targeted few-shot fine-grained image classification methods.These methods are based on the learning framework of few-shot and mine the classification features of images while minimizing dependence on data volume.However,most of these methods overlook the relationship between the features of the support set and the query set,resulting in unsatisfactory classification performance.This paper proposes two new methods for few-shot fine-grained image classification based on the local features and the relationship between local information of support set and query set samples.The main research content is as follows:(1)A Few-shot fine grained image classification method based on local attention fusion is proposed.The core idea of the method is to enhance the learning ability of local discriminative features in the model,thereby enhancing the classification ability.Specifically,the designed local attention generation module generates local attention maps by extracting pixel level features,and fuses the generated local attention maps with advanced global Semantic information to generate identification area features to guide model classification,so that the model pays more attention to the key local areas of query sets and support sets.Finally,the similarity measurement module is used to calculate the distance from the local features of the query set to the support set category,completing the classification.The model was validated under few-shot tasks on three classic finegrained datasets.The experimental results show that the classification performance of the method is improved compared to other related research methods.The ablation experiment also verifies the effectiveness of each module in the model.Visual analysis also shows that the discriminative attention generation module can indeed focus on the discriminative regions of the image.(2)A Few-shot fine grained image classification method based on local information cross is proposed.The core idea of the method is to establish a semantic relationship for local information exchange between the support set and query set samples.Specifically,the double attention feature encoder is designed to extract the deep information with strong semantic information and the shallow information with location and structure information,and further fuse and extract key local information.Then,through the interaction of support set and query set images on the semantic information of local features,the feature representation ability is improved,the local semantic consistent correlation of support set and query set is established,and the semantic inconsistent features are filtered out to obtain reliable common concerns,improve the image prototype feature representation ability,and improve the classification accuracy of the model.Finally,reconstruct the loss function to calculate the loss.The experimental results on three classic fine-grained datasets show that the method has achieved competitive results compared to existing mainstream methods,and ablation experiments and visualization analysis have also verified the impact of each module of the method on the model.