Prompt-Driven Multitask Learning With Task Tokens for ORSI Salient Object Detection

Salient object detection in optical remote sensing images (ORSI-SOD) is challenging due to complex backgrounds and subtle boundary cues. While prior methods often introduce edge-guided supervision or auxiliary branches to refine boundaries, they commonly suffer from task interference and inefficient optimization. To address this, we propose the Multitask Prompt Network (MTPNet), a prompt-driven framework that incorporates explicit task conditioning through spatial task prompts and task-specific tokens. Task prompts are lightweight, learnable priors injected into both the encoder and decoder to guide task-aware feature learning. In contrast, task tokens are introduced in the decoder to serve as compact semantic queries that selectively aggregate task-relevant information and enable token-level task communication. A task prompt discrimination loss is introduced to promote representation disentanglement and mitigate intertask redundancy. In addition, a Progressive Linear Attention module is designed for efficient global context modeling, and a Saliency Edge-Aware Alignment Loss is employed to improve boundary precision. Extensive experiments on three ORSI-SOD benchmarks show that MTPNet outperforms 19 state-of-the-art methods while maintaining high efficiency.