Single Image Reflection Removal via inter-layer Complementarity.

Yue Huang, Zi'ang Li, Tianle Hu, Jie Wen, Guanbin Li, Jinglin Zhang, Guoxu Zhou, Xiaozhao FangarXiv 2025 · OpenReview

Abstract

Single Image Reflection Removal (SIRR) is a fundamentally ill-posed problem, often described as the "cocktail party problem" of computer vision. Traditional deep learning methods struggle to balance fine-grained detail restoration with computational efficiency, often relying on massive models or coarse-grained priors. This paper introduces the Dual-Prior Interaction Transformer (DPIT), which utilizes a novel Local Linear Correction Network (LLCN) to generate high-quality transmission priors. By shifting from direct pixel generation to a local linear correction strategy ($T = sI + b$), DPIT preserves original image textures while significantly reducing parameter overhead. Integrated with a Dual-Stream Channel Reorganization Transformer, the model achieves state-of-the-art results across multiple real-world benchmarks with approximately 40% of the parameters of previous leading models.

Introduction

When capturing images through glass, the resulting photograph $I$ is a superposition of the background transmission $T$ and the reflected scene $R$, modeled as $I = T + R$. Separating these components from a single image is a classic "ill-posed" challenge because infinite combinations of $T$ and $R$ can produce the same $I$.

Recent advancements have leveraged "priors"—additional information such as semantic hints or coarse reflection estimates—to guide the separation process. However, high-performance models that generate fine-grained priors are typically computationally prohibitive, while lightweight alternatives often lack the accuracy needed for high-quality restoration. DPIT addresses this gap by providing high-quality, fine-grained guidance without the massive computational tax associated with traditional generative architectures.

Methodology

The DPIT architecture is built upon two primary innovations designed to maximize efficiency and restoration quality: the Local Linear Correction Network (LLCN) and the Dual-Stream Channel Reorganization Transformer.

Local Linear Correction Network (LLCN)

Instead of attempting to "hallucinate" transmission pixels from scratch, the LLCN adopts a "selection rather than generation" strategy. It utilizes a physical model where the transmission is estimated as $T = sI + b$. The network predicts per-pixel scaling ($s$) and bias ($b$) factors. This approach ensures that the background information already present in the mixed image is preserved and enhanced rather than replaced, allowing the model to achieve high accuracy with a significantly lower parameter budget.

Dual-Stream Channel Reorganization Transformer (DSCRT)

To effectively fuse the LLCN's transmission prior with general semantic features (extracted via a pre-trained Swin Transformer), the authors developed a specialized attention mechanism within the DSCRA blocks. Standard dual-stream transformers often suffer from high computational overhead. The proposed mechanism reorganizes channels so that both processing streams contain a mix of heterogeneous features. This allows the attention layers to perform intra-stream separation and cross-stream complementation simultaneously and efficiently.

Overall architecture of the Dual-Prior Interaction Transformer (DPIT), highlighting the Local Linear Correction Network and the interactive dual-prior feature extraction streams.

Experiments and Results

Datasets and Training

The model was trained on a hybrid dataset comprising synthetic pairs (generated from PASCAL VOC using non-linear blending) and 289 real-world pairs. Evaluation was conducted across five diverse real-world benchmarks: Real20, Objects, Postcard, Wild, and Nature, totaling 494 images.

Quantitative Performance

DPIT achieved a new state-of-the-art (SOTA) average PSNR of 27.21 dB and an SSIM of 0.924 across the test sets. Notably, it outperformed the previous leading model, RDNet, by 0.49 dB in PSNR while using only 131.54M parameters—roughly 41.6% of RDNet's 315.89M parameters.

Quantitative comparison of DPIT against state-of-the-art methods, demonstrating superior performance (PSNR/SSIM) and significant parameter efficiency.

Qualitative Analysis

Visual results demonstrate that DPIT is particularly effective at removing complex "ghosting" effects and preserving fine structural details in areas of high reflection concentration, such as bridge railings or dark indoor corners. Ablation studies confirmed the importance of the LLCN; removing the transmission prior resulted in a performance drop of over 1 dB in PSNR.

Conclusion

DPIT represents a significant shift in SIRR research by demonstrating that "selection" via local linear correction is more efficient and effective than direct pixel generation. By leveraging inter-layer complementarity through channel reorganization, the model provides a high-performance, lightweight solution for real-world reflection removal. Future work may explore the integration of additional priors, such as depth maps or polarization data, into this efficient transformer framework.

Summary updated

BibTeX
@misc{huang2025single,
  author = {Yue Huang and Zi'ang Li and Tianle Hu and Jie Wen and Guanbin Li and Jinglin Zhang and Guoxu Zhou and Xiaozhao Fang},
  title = {{Single Image Reflection Removal via inter-layer Complementarity.}},
  year = {2025},
  url = {https://openreview.net/forum?id=QxsXxyR4im},
}
Figure