ARConvL: Adaptive Region-Based Convolutional Learning for Multi-class Imbalance Classification

Real-world image classification usually suffers from the multi-class imbalance issue, probably causing unsatisfactory performance, especially on minority classes. A typical way to address such problem is to adjust the loss function of deep networks by making use of class imbalance ratios. However, such static between-class imbalance ratios cannot monitor the changing latent feature distributions that are continuously learned by the deep network throughout training epochs, potentially failing in helping the loss function adapt to the latest class imbalance status of the current training epoch. To address this issue, we propose an adaptive loss to monitor the evolving learning of latent feature distributions. Specifically, the class-wise feature distribution is derived based on the region loss with the objective of accommodating feature points of this class. The multi-class imbalance issue can then be addressed based on the derived class regions from two perspectives: first, an adaptive distribution loss is proposed to optimize class-wise latent feature distributions where different classes would converge within the regions of a similar size, directly tackling the multi-class imbalance problem; second, an adaptive margin is proposed to incorporate with the cross-entropy loss to enlarge the between-class discrimination, further alleviating the class imbalance issue. An adaptive region-based convolutional learning method is ultimately produced based on the adaptive distribution loss and the adaptive margin cross-entropy loss. Experimental results based on public image sets demonstrate the effectiveness and robustness of our approach in dealing with varying levels of multi-class imbalance issues.