Research On Combing Foreground Feature Boosting And Diversification Feature For Fine-grained Image Classification

Fine-grained image classification is a challenging task due to deep feature extraction is interrupted by sophisticated background noises,the classification network often focuses on the global sematic feature while neglecting other inconspicuous but distinguishable region parts,and they treat different part features in isolation while neglecting their relationships.To handle these limitations,fine-grained image classification for combing foreground feature boosting and diversification feature learning is proposed.Firstly,in order to eliminate the influence of background noises on feature extraction,a foreground feature boosting module is designed to capture spatial correlation between each two pixels from the global field,ensure the consistency within the class,and improve confidence level positioning of foreground object boundary box,so as to achieve foreground feature amplification.Secondly,in order to solve that the classification network often focuses on the global sematic feature while neglecting other inconspicuous but distinguishable region parts,contradictory feature modules are embedded in different layers of the network to obtain the high response feature information of each layer,and suppresses it to force the following network to mine other potential parts.Then,in order to enrich the semantic complementary information in each region part feature,a region feature complementary module is designed to enhance the feature representation and enrich the semantic information by integrating other information through negative matrix multiplication;Finally,our classification model uses cross entropy loss to constraint the network.In the quantitative verification of the proposed method,accuracy is taken as evaluation index,The experimental results show that: on the two public fine-grained image data sets CUB-200-2011 and FGVC-Aircraft,the proposed classification model gets good results,with the accuracy reaching 88.97% and 93.94%,respectively.Compared with other fine-grained image classification methods,the proposed method gets excellent performance in accuracy.There are 32 pictures,8 tables and 59 references.