WMul-Net: A Weakly Supervised Multi-Branch Residual Network for Fine-Grained Classification

Weakly supervised fine-grained image classification methods have gained significant attention due to their reliance solely on image-level labels, avoiding the need for additional annotations such as bounding boxes or keypoints. Despite this, fine-grained feature extraction still requires further improvement. In this paper, we propose WMul-Net, a weakly supervised approach for fine-grained image classification. WMul-Net leverages multi-branch residual blocks that extract features at different scales, preserving feature diversity to reduce intra-class variations and inter-class similarities. Additionally, our network distills correlations between class labels and background information, enhancing fine-grained image recognition. Experimental results demonstrate that WMul-Net outperforms existing methods on the Stanford Cars and FGVC-Aircraft datasets, achieving accuracies of 95.64% and 94.5%, respectively.