Fine-grained Image Classification Based On Adaptive Progressive Sampling And Few-shot Learning

With the growing demand for smart technology,traditional image classification methods based on big data and cross-semantics have been difficult to meet the needs of real scenes.Among them,in the task of image classification,fine-grained image classification has gradually become a research hotspot due to the problems of "large intra-class difference,small inter-class difference" and difficulty in forming a large data set with uniform distribution.On the one hand,existing fine-grained image classification researches focus on exploring new feature fusion methods in order to mine more subtle discriminant features.On the other hand,in order to overcome the challenge of limited sample size,researchers adopt the few-shot learning method in order to obtain better classification performance under limited data samples.This approach improves the generalization ability of the model by learning abstract concepts and patterns from a small number of samples,and achieves higher accuracy in practical applications.This paper aims to conduct research on fine-grained image classification,mainly from two aspects: On the one hand,it explores how to solve the challenges faced by traditional cross-semantic and largedata image classification methods in fine-grained image classification;On the other hand,by using few-shot learning method,trying to obtain higher classification performance under limited data samples.These studies aim to improve the accuracy and engineering practicability of finegrained image classification.The paper mainly includes the following three parts:(1)A multi branch feature learning approach for fine-grained visual recognition is proposed.The method has multiple classification branches,each branch can extract different granularity of feature information,and forms an adaptive progressive sampling mode.In order to enhance the feature representation ability,a bottom-up down-sampling feature fusion method is introduced,which can effectively combine the high-level features with the low-level features.(2)A cross-category information fusion method for few-shot fine-grained recognition is proposed.This method uses the features of the query sample and different support sample class cluster features to get relation maps,which is used to improve the feature expression ability of query samples;at the same time,the query sample feature adjusts channel weights by selflearning within channels;at last,local feature metric is used to classified for query samples.According to the above method,the model can pay more attention to the target characteristics of the query samples,so as to improve the classification performance.(3)A foreground and background separation measurement method for few-shot fine-grained recognition is proposed.Based on the observation of feature descriptors,the concept of foreground feature descriptors and background feature descriptors is introduced,and a new local feature measurement method is applied.In addition,the global features of support samples are used to generate the attention map,which is also based on the metric method.The attention map is used to modify the weights of different feature descriptors of the original query samples,so that minimizes the background error interference to the similarity score.