Bi-Branch Multiscale Feature Joint Network for ORSI Salient Object Detection in Adverse Weather Conditions

• Published in: IEEE transactions on geoscience and remote sensing 2024, Vol.62, p.1-10
• Main Authors: Yuan, Jianjun, Zou, Xu, Xia, Haobo, Liu, Tong, Wu, Fujun
• Format: Article
• Language: English
• Subjects: Adverse weather conditions; Aggregation; bi-branch joint structure; Components; Deep learning; Feature extraction; Image enhancement; Meteorology; Modules; multiscale feature; Object detection; Object recognition; Optical imaging; optical remote sensing images (ORSIs); Optical sensors; Rain; Remote sensing; Robustness; Salience; salient object detection (SOD); Semantics; Synthetic data; Weather
• ISSN: 0196-2892; 1558-0644
• DOI: 10.1109/TGRS.2024.3485586

Salient object detection (SOD) of optical remote sensing images (ORSIs) has been a crucial part of the remote sensing field. In recent years, with the development of deep learning, many salient detection models for ORSIs have emerged. However, current study is limited to sunny weather conditions, and there is a lack of research on SOD in adverse weather conditions. Traditional models lack robustness and tend to miss detection in adverse weather conditions. To address this challenge, this article proposes a bi-branch multiscale feature joint network (BMFJNet) that achieves SOD in adverse weather conditions through a bi-branch linear joint structure. First, we obtain clean ORSIs through the dark channel prior and feed the clean images and the hazy images by two linear branches to the backbone for feature extraction, respectively. Second, the obtained effective features are input to the detection module for salient analysis. The detection module consists of three key components, where the multiscale feature aggregation module (MFAM) achieves salient feature enhancement in both dimensional directions through an attention mechanism, the adjacent pooling guidance module (APGM) guides the contextual information of adjacent layers through multiple pooling layers, and the feature fusion module aggregates global information from different components. In addition, we introduce a self-supervised robust restoration loss that enables our network to cope with different levels of adverse weather. Extensive experiments on synthetic datasets demonstrate the superiority of our proposed model over other state-of-the-art models on various metrics.