Channel Attention Networks

Abstract

Multi-band images beyond RGB are becoming popular in both commercial applications and research datasets, yet existing deep learning models were designed for academic RGB datasets. In this talk, we propose Channel Attention Networks (CAN), a deep learning model that uses soft attention on individual channels. We jointly train this model end-to-end on Spacenet, a challenging multi-spectral semantic segmentation dataset. In a comparative study, CAN outperforms previous models. We also demonstrate that CAN is significantly more robust to noise in individual bands than the other models, because the attention network allocates attention away from the noisy channels. Our proposed method marks the first step in designing deep learning algorithms specifically for multi-spectral imagery. Semantic Segmentation; Convolutional Neural Networks; Attention