Weakly Supervised Fine-Grained Visual Classification Through Spatial Information Mining and Attention-guided Regularization

Over-fitting is a severe problem when we adopt deep neural networks with a large number parameters in fine-grained visual classification. Many data augmentation methods are proposed through weakly supervised learning to alleviate over-fitting issue. Different from those methods, we propose a weakly supervised attention-guided regularization by object parts’ attention maps to fine-tune the Fully Connected (FC) layer and relieve over-fitting issue during training in this paper. On the other hand, the neural units in the last convolutional layer contain the same receptive fields that limit recognition performance due to involving lots of background noises. To alleviate this issue, we devise a spatial information mining module with an auxiliary penalty loss to aggregate multi-scale receptive fields feature maps with the selected precedent layer. Comprehensive experiments are conducted to show our method achieves or surpasses state-of-the-art results on common fine-grained classification datasets.