A Discriminative Feature Extraction Network For Fine-grained Image Classification

In recent years,in the context of the era of big data,digital multimedia technolo-gy has developed rapidly,and massive image data have flooded into the Internet.It is particularly important to automatically classify and identify relevant information with the help of computer.With the upsurge of deep learning,the development of image recognition has ushered in new vitality.The traditional coarse-grained image classifi-cation is difficult to meet the recognition needs of users in real life.The fine-grained classification of images has gradually attracted the attention of researchers.Fine-grained image classification remains a challenging problem due to the subtle inter-class differences and large intra-class variances.Currently,most existing methods tend to solve the problem by first discovering subtle differences as well as discrimina-tive parts,and then extracting features from them for final classification.Although this strategy has been proved to be effective,there are still some critical problems.1)It is hard to guarantee that the data-driven weakly supervised network could find subtle and discriminative regions accurately.Generally speaking,it is a weakly-supervised detection problem,which has not been well solved.Besides,2)valuable information may be distributed over different parts of an image,how to discover all these parts in one model efficiently and completely is still an open question.What's more,3)while most existing fine-grained image classification methods only concentrate on subtle re-gions and neglect other useful information,and the information of different levels can not be effectively utilized.Inaccuracy of localization and incompleteness of valuable information caused by these problems may affect the final performance.Therefore,this thesis proposes a Style-Guided discriminative feature Extraction Network,SGEN for short,to tackle these problems from a different point of view.The feature extraction network is proposed to extract useful information from different as-pects of an image without the necessity of discovering discriminative parts firstly.More specifically,it is trained with an encoder-decoder architecture to learn the mapping from images to a latent space;thereafter,the extracted latent vectors are utilized to reveal the discriminative features from different levels that are critical for fine-grained classifica-tion.To describe the target image more comprehensively,a complementary network is combined to further provide global features.Finally,an improved loss function is designed for fine-grained image classification,which focuses on distinguishing class-es that are easily confused and reducing over-fitting.Extensive experiments on widely benchmark datasets demonstrate that SGEN is able to achieve promising performance and outperform some existing state-of-the-art methods.In summary,the main contribu-tions of this thesis include:?A style-guided discriminative feature extraction network is introduced.Accord-ing to the investigation,SGEN is the first fine-grained image classification method that tries to extract different level features to describe different aspects of an im-age with the guide of a StyleGAN and the latent space defined by it.?An improved loss function is designed for fine-grained image classification,which aims to distinguish the correct class from easily confused ones and reduce over-fitting.?Extensive experiments are conducted on three widely-used datasets:CUB-200-2011,Stanford Cars and FGVC-Aircraft,and the results demonstrate the superi-ority of SGEN.