Title: Principled Reflection Separation via Nonlinear Superposition and Feature Interaction

Abstract:

The separation of reflections from a single image is inherently difficult due to the entanglement of transmission and reflection components within complex image formation processes. Current methods often depend on simplified assumptions or treat layers independently, which restricts their effectiveness in practical, real-world applications. This study re-examines the problem through a unified lens, highlighting a critical flaw in prevailing techniques: the standard linear composition model used in the sRGB domain cannot adequately represent the nonlinear coupling generated by real-world image signal processing pipelines.

To overcome this limitation, we propose a learnable nonlinear superposition model that more accurately depicts layer interactions, thereby enhancing the fidelity of the decomposition. Leveraging this formulation, we introduce a generalized dual-stream interactive framework designed to explicitly capture the bidirectional dependencies between the transmission and reflection layers via feature exchange. This framework integrates activation-, gating-, and attention-based interaction mechanisms and remains compatible with both CNN and Transformer architectures.

Comprehensive experiments across various real-world benchmarks show that our method delivers superior performance and robust generalization. Crucially, our findings suggest that reflection separation should not be viewed as merely reversing a linear mixture; rather, it involves learning the underlying nonlinear formation and interaction processes. This perspective provides fresh insights for designing principled image decomposition models. Code and models are accessible at https://mingcv.github.io/DIRS-Page.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC