Combining Multi-Feature Regions for FineGrained Image Recognition
Sun Fayou, Hea Choon Ngo, Yong Wee Sek
Abstract
Fine-grained visual classification(FGVC) is challenging task duo to the subtle discriminative features.Recently, RA-CNN selects a single feature region of the image, and recursively learns the discriminative features. However, RA-CNN abandons most of feature regions, which is not only the inefficient but aslo ineffective.To address above issues,we design a noval fine-grained visual recognition model MRA-CNN,which associates multi-feature regions.To improve the feature representation,attention blocks are integrated into the backbone to reinforce significant features;To improve the classification accuracy, we design the feature scale dependent(FSD) algorithm to select the optimal outputs as the classifier inputs;To avoid missing features, we adopt the k-means algorithm to select multiple feature regions.We demonstrate the value of MRA-CNN by expensive experiments on three popular fine-grained benchmarks:CUB-200-2011,Cars196 and Aircrafts100 where we achieve state-of-the-art performance.Our codes can be found at https://github.com/dlearing/MRA-CNN.git.