A Spatial-Channel Feature-Enriched Module Based on Multicontext Statistics Attention
Huanjie Tao, Qianyue Duan
Abstract
Convolutional neural networks (CNNs) have demonstrated remarkable performance in various computer vision tasks, such as image classification, semantic segmentation, and object detection. However, learning discriminative and generalizable features that can overcome both intra-class variations and inter-class ambiguity remains a challenging problem. To address this issue, we propose a novel module called the spatial-channel feature-enriched module (SCFEM), which can be easily integrated into existing CNN architectures. Specifically, we propose a refined multi-scale residual block (RMRB) to dynamically produce channel-wise weights via a shared aggregation gate to selectively fuse multi-scale features. Moreover, a multi-context statistics attention block (MSAB) is proposed to explore both the semantic dependency between channels and the long-range spatial contextual dependency between pixels via multi-context channel attention (MCA) and statistic spatial attention (SSA). MCA captures the local context information of each channel by considering the associated neighbor channels. SSA captures complex long-range dependencies and discriminative but subtle differences among pixels by leveraging high-order statistics. Experimental results demonstrate the superiority of our SCFEM over state-of-the-art methods on multiple computer vision tasks.