ANNet: Asymmetric Nested Network for Real-Time Cloud Detection in Remote Sensing

Cloud detection is one of the crucial tasks in the field of remote sensing, which is regarded as binary classification at the pixel level. Although a large number of recent deep learning-based methods have made great progress, most of their appealing performances come at the expense of a large amount of computation, which reduces the real-time performance accordingly. In order to bridge the gap between segmentation performance and inference speed, we propose a novel asymmetric nested network (ANNet) architecture termed ANNet, which is designed for real-time cloud detection with excellent performance. In the encoder branch of ANNet, we introduce an effective tiny U-shape block (TUB) to enrich detailed spatial contexts in each stage, which also allows ANNet to embed spatial recovery ability earlier, and a lightweight, simple feature fusing module (SFFM) is designed to refine the semantic features map at a lower level of TUB for better performance. Following the encoder, which is diverse from the most symmetric U-shape approaches, an asymmetric and lightweight decoder (ALD) with only convolution and bilinear up-sample operations is employed for spatial recovery. We also, moreover, demonstrated that using a constant channel size instead of a larger channel volume as the network goes deeper is an efficient and effective design for cloud detection tasks. Substantial experiments are performed on GF1_WHU and 95-Cloud datasets, which show that ANNet has achieved excellent performance with low computation cost compared to most of the existing state-of-the-art methods. On GF1_WHU dataset, ANNet-l achieves 93.79% on mIoU at 125 FPS, while ANNet-s with only 29.5 K parameters yields 90.74% mIoU at 251 FPS on the Nvidia RTX 2080Ti.