A Cross-Fusion Network for Salient Object Detection in Optical Remote Sensing Images

Abstract

Salient object detection (SOD) is usually taken as a key procedure on the preprocessing of remote sensing images (RSIs), in which RSI-SOD techniques are employed to accurately locate the most attractive targets from RSIs. The existed RSI-SOD models, however, face a challenge on how to balance global context and local detailed information efficiently due to varying object scales and cluttered backgrounds in RSIs. Also to improve the portability of the network to facilitate the practical application of the model, we propose a efficient network, multieffective combined network (MECNet). MECNet combines multiscale networks with an edge detection auxiliary network, utilizing an attention mechanism for enhanced performance. Within MECNet, the multiview combination block employs an attention-based approach to capture rich contextual information across scales, aiding in the detection of various-sized objects. The post-aggregation reassignment block utilizes multiscale fusion and edge features generated by the edge detection network to enrich semantic and detailed information, effectively handling intricate details. The channel enhancement decoder module employs channel attention to amplify channel cues, enhancing the detail quality of the prediction maps. Evaluated against state-of-the-art methods, MECNet demonstrates superior performance making it a promising solution for practical RSI-SOD applications.