Improving Residual Block for Semantic Image Segmentation

Currently, most state-of-the-art semantic segmentation methods employ residual network as base network. Residual network is composed of residual blocks. In this paper, we present an improved residual block called pyramid residual block to explicitly exploit context information and enhance useful features. In contrast to the standard residual block, the proposed pyramid residual block contains two newly added components: pyramid pooling module and attention mechanism. The former aggregates different-region-based context information. And the latter is able to adaptively re-calibrate feature responses through element-wise multiplication operation, thus enhancing useful features and suppressing less useful ones. Our proposed pyramid residual block demonstrates outstanding performance in PASCAL VOC 2012 segmentation datasets, and improve the segmentation accuracy by a large margin over the standard residual block.