Multi-Scale Weighted Branch Network for Remote Sensing Image Classification

Abstract

Remote sensing image classification aims to assign semantic label for each input pixel. In this paper, we propose a Multi-Scale Weighted Branch Network (MSWBN) for this dense prediction task. Inspired by attention module, which is commonly adopted to enhance the informative features among the dense feature maps in the deep network, we firstly introduce a Hierarchical Weighted branch Module (HWM). The HWM is designed to extract multi-scale information from the backbone simultaneously with a weighted branches architecture, whose branch weights are generated from lower layers of the backbone. Then, a Low level features Branch Module (LBM) is proposed to embed information with high resolution, where the weighted sum of output from the HWM and low level features is calculated as the dense prediction of the proposed Multi-Scale Weighted Branch Network. The proposed method outperforms extisting best models on the large scale remote sensing image classification dataset (GID) in terms of both efficiency and accuracy.