PhaSR: Generalized Image Shadow Removal with Physically Aligned Priors

2026-01-24

AI Summary

Existing shadow removal methods suffer from limited performance in multi-light-source environments due to misalignment between physical priors and data-driven features. This work proposes a dual-level prior alignment framework that first employs Physically Aligned Normalization (PAN) to achieve closed-form illumination correction, followed by a Geometry-Semantic Rectification Attention (GSRA) mechanism that fuses depth geometry with DINO-v2 semantic features to enhance cross-modal consistency. For the first time, this approach enables synergistic alignment between physical priors and semantic-geometric representations, significantly improving robustness and generalization across scenarios ranging from single to complex multi-light settings. The method achieves superior performance on multiple real-world shadow datasets while maintaining lower computational complexity compared to existing approaches.

Technology Category

• Computer Vision: Low Level & Physics-based Vision
• Intelligent Robots: Multimodal Perception & Sensor Fusion
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Security and Privacy: Data transparency and provenance
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Shadow removal under diverse lighting conditions requires disentangling illumination from intrinsic reflectance, a challenge compounded when physical priors are not properly aligned. We propose PhaSR (Physically Aligned Shadow Removal), addressing this through dual-level prior alignment to enable robust performance from single-light shadows to multi-source ambient lighting. First, Physically Aligned Normalization (PAN) performs closed-form illumination correction via Gray-world normalization, log-domain Retinex decomposition, and dynamic range recombination, suppressing chromatic bias. Second, Geometric-Semantic Rectification Attention (GSRA) extends differential attention to cross-modal alignment, harmonizing depth-derived geometry with DINO-v2 semantic embeddings to resolve modal conflicts under varying illumination. Experiments show competitive performance in shadow removal with lower complexity and generalization to ambient lighting where traditional methods fail under multi-source illumination. Our source code is available at https://github.com/ming053l/PhaSR.

Problem

• shadow removal
• illumination disentanglement
• physically aligned priors
• multi-source lighting
• intrinsic reflectance

Innovation

• Physically Aligned Normalization
• Geometric-Semantic Rectification Attention
• Shadow Removal
• Cross-modal Alignment
• Retinex Decomposition