INP-Net: Implicit Neural Prompting Network for Remote Sensing Image Dehazing

Restoring high-quality images from hazy observations is crucial for visual perception and downstream detection in remote sensing applications. Recent deep learning-based methods have achieved remarkable progress in dehazing, however, they often suffer from either unreliable prompting guidance or inadequate global modeling. To bridge these gaps, we propose INP-Net, an Implicit Neural Prompting Network for remote sensing image dehazing. INP-Net introduces an implicit neural prompting mechanism that exploits learnable implicit neural representations (INR) in the YCbCr color space as degradation-insensitive prompts, which are dynamically injected into the RGB decoding pipeline. In addition, we design the Adaptive Sampling and Prior-enhanced (ASP) Transformer Block to enrich global feature diversity. The ASP comprises two core components: Adaptive Soft Sampling Self-Attention (ASSA), which performs probabilistic token selection to eliminate channel redundancy and enhance discriminative representations; and the Prior-Modulated Mixed-Scale Feed-Forward Network (PMFN), which leverages haze prior knowledge as a multi-scale modulator to guide local feature restoration. Extensive experiments on multiple benchmark datasets demonstrate that the proposed INP-Net surpasses state-of-the-art methods both quantitatively and qualitatively. Our method achieves PSNR gains of 0.47 dB and 0.75 dB over state-of-the-art methods on the challenging StateHaze1K-thick and NH-Haze datasets, respectively.