Ori-INR: Orientation-Aware Implicit Neural Representation for Remote Sensing Image Representation

Implicit Neural Representations (INRs) model images as continuous functions that map spatial coordinates to pixel values, providing a compact and resolution-independent framework for image representation. Existing INR methods have achieved strong performance through coordinate-based modeling and activation design, enabling effective representation of complex visual signals. Remote sensing images provide a representative scenario where spatial orientation plays an important role in visual representation. Many structures, such as roads, rivers, and building boundaries, exhibit clear geometric layouts and orientation-dependent patterns, making orientation-aware modeling particularly important for accurate representation and interpretation. In this work, we propose Ori-INR, an orientation-aware implicit neural representation for remote sensing image representation. Ori-INR introduces explicit orientation modeling into coordinate-based networks from two complementary perspectives. First, we design an orientation-modulated activation that enables hidden representations to capture orientation-sensitive responses aligned with spatial structures. Second, we introduce a radial-angular coordinate encoding that integrates radial distance and angular phase information, providing explicit orientation-aware cues at the input level. By combining these two components, Ori-INR enhances the ability of INRs to represent orientation-structured visual signals in a more expressive and structured manner. Extensive experiments on multiple remote sensing benchmarks demonstrate that the proposed method consistently improves representation quality and better preserves geometric and directional structures compared with existing INR approaches.