Efficient Salient Object Detection in Optical Remote-Sensing Imagery through Semantic Matching and Edge Alignment

Abstract

Recent advancements in salient object detection (SOD) for optical remote-sensing images (ORSI) have largely relied on convolutional neural networks (CNNs). However, these methods often overlook the high number of parameters and computational costs associated with CNNs, with only a few focusing on portability and efficiency. To address this gap and enhance practical applications, we introduce a novel lightweight network for ORSI- SOD called SeaNet. SeaNet incorporates several key components: a lightweight MobileNet-V2 for feature extraction, a dynamic semantic matching module (DSMM) for high-level feature processing, an edge self-alignment module (ESAM) for low-level feature refinement, and a portable decoder for inference. The process begins by compressing high-level features into semantic kernels. These kernels then activate salient object locations in two groups of high-level features through dynamic convolution operations within the DSMM. Concurrently, the ESAM self-aligns cross-scale edge information from two groups of low-level features using L2 loss, enhancing detail. The decoder subsequently infers salient objects from the highest-level features, leveraging the accurate locations and fine details provided by the modules.Extensive experiments on two public datasets demonstrate that SeaNet not only surpasses most state-of-the-art lightweight methods in performance but also achieves comparable accuracy to conventional state-of-the-art methods. Remarkably, it does so with only 2.76 million parameters and a computational cost of 1.7 billion floating-point operations (FLOPs) for 288x288 inputs.