Adaptively Weighted Copy-Decoupling Resampling Strategy for Long-Tailed Multi-label Classification

Typical natural images are associated with multiple classes, and different classes typically exhibit a long tail distribution. The Binary Relevance strategy is commonly employed to deal with multi-label image classification. However, this strategy faces two major challenges: the resampling strategy’s failure due to label co-occurrence and the classifier’s bias towards negative samples with imbalanced properties of positive and negative samples. To overcome these limitations, this paper proposes two novel approaches: a simple but efficient sampling strategy, i.e., Copy-Decoupling Resampling (CD-RS), and an adaptive weighted loss function, i.e., Adaptive Focal Loss (AFL). Specifically, CD-RS converts a multi-label image into multiple single-label images with special labels, eliminating the effect of label co-occurrence on the resampling strategy. Meanwhile, AFL adaptively adjusts the focal factor in focal loss according to model performance on samples, which gives more appropriate weights to samples and alleviates the bias caused by sample imbalance during classifier training. Finally, this paper proposed method achieved impressive results on popular datasets VOC-MLT and COCO-MLT, with mAP of 78.96 and 55.35, respectively.