Abstract
Fine-grained visual categorization is a special case in image classification. It is a challenging task in which objects may have small between-class variation and large intra-class variation caused by viewpoints, pose and lighting condition changes. In order to improve the performance of classification, we incorporate the part information of objects and propose a part-based classification framework for fine-grained visual categorization. The proposed classification framework consists of the following steps: First, we infer the part segmentation from foreground regions and part locations of the object. With the inferred part segmentation, we implicitly perform pose normalization on the object. Then, we extract features from the corresponding part segments and apply feature encoding to generate the final image representation. Finally, we perform image classification based on their collaborative representation with regularized least squares from the whole training data.