ReflexSplit: Single Image Reflection Separation via Layer Fusion-Separation

1. Problem Statement

Single Image Reflection Separation (SIRS) aims to disentangle a mixed image $I$ into a transmission layer $T$ (the background scene) and a reflection layer $R$ (the unwanted reflection).

• Core Challenge: Existing methods often suffer from transmission-reflection confusion, particularly in deep decoder layers. This occurs because implicit fusion mechanisms and inadequate multi-scale coordination cause features from the two layers to become entangled.
• Limitations of Prior Work:
  • Linear Models: Early models ($I = T + R$) fail to capture layer asymmetry.
  • Nonlinear Models: While recent methods use learnable nonlinear functions ($I = T + R + \Phi(T, R)$), they often lack explicit mechanisms to separate layers during the decoding process.
  • Deep Network Issues: Increasing network depth often leads to information loss and feature entanglement, where the network struggles to distinguish between shared structures and layer-specific details, resulting in artifacts like blurred boundaries or residual reflections.

2. Methodology: ReflexSplit

ReflexSplit is a dual-stream framework designed to explicitly alternate between fusion (for shared structure extraction) and separation (for layer-specific disentanglement). It consists of three core components:

A. Dual-Branch Feature Extraction

To capture complementary representations, the encoder uses two parallel branches:

1. Global Feature Extractor Block (GFEB): A pre-trained Swin Transformer that extracts semantic priors ($P_\ell$) at multiple scales.
2. Local Feature Extractor Block (LFEB): A MuGI-based CNN that captures fine-grained texture details ($E_\ell$) at hierarchical resolutions.
   This design prevents the feature entanglement common in single-stream encoders.

B. Cross-scale Gated Fusion (CrGF)

Located at decoder levels $\{4, 3, 2\}$, CrGF addresses gradient instability and feature degradation by adaptively aggregating features from heterogeneous sources (semantic priors, texture details, and decoder context).

• Mechanism: It uses bidirectional gating to balance current-level features with context from deeper layers.
• Operation: It computes a main path ($F_{main}$) and an auxiliary path ($F_{aux}$) using gating functions ($G_1, G_2$) to select complementary channels, followed by a weighted fusion.
• Goal: To stabilize gradient flow and maintain hierarchical consistency, preventing the progressive confusion of transmission and reflection features.

C. Layer Fusion-Separation Block (LFSB)

This is the core innovation for explicit disentanglement. LFSB alternates between two operations:

1. Early Fusion (Bidirectional Projection): Aligns the semantic spaces of the transmission and reflection streams via cross-stream projection ($W_t, W_r$). This allows streams to share complementary information before separation.
2. Differential Dual-Dimensional Attention: Inspired by the Differential Transformer, this module performs cross-stream subtraction to suppress interference.
   • It computes Self-Attention (SA) and Cross-Attention (CA) maps for both streams.
   • Differential Operator: Instead of simple aggregation, it applies:
     $$A^t_{diff} = (A^t_{SA} + A^t_{CA}) - \sigma(\lambda_\ell)(A^r_{SA} + A^r_{CA})$$
     $$A^r_{diff} = (A^r_{SA} + A^r_{CA}) - \sigma(\lambda_\ell)(A^t_{SA} + A^t_{CA})$$
   • Here, $\lambda_\ell$ is a learnable parameter that controls the strength of suppression. This explicitly forces the network to subtract the "other" layer's features, ensuring distinct representations.

D. Curriculum Training Strategy

To prevent the differential separation from destabilizing early training (when features are poorly structured), the authors employ a curriculum learning approach:

• Depth-Dependent Initialization: Deeper layers (global) start with stronger separation weights ($\lambda \to 0.8$), while shallow layers (local details) start with weaker weights ($\lambda \to 0.2$).
• Epoch-Wise Warmup: The global separation strength $\lambda_{diff}$ is linearly increased from 0.1 to 1.0 over the first 30 epochs.
• Effect: The network first learns holistic reconstruction and shared structures, then progressively focuses on layer-specific disentanglement.

3. Key Contributions

1. Explicit Layer Fusion-Separation Framework: A paradigm that alternates between fusion (for shared degradation) and differential separation (for disentanglement), effectively preventing transmission-reflection confusion throughout the network hierarchy.
2. Cross-scale Gated Fusion (CrGF) & Differential Attention: CrGF ensures multi-scale consistency, while the LFSB uses cross-stream attention subtraction to explicitly disentangle layers, a mechanism not present in previous SIRS methods.
3. Curriculum Training: A novel training schedule that stabilizes the learning of differential separation, allowing the model to handle complex nonlinear mixing without feature collapse.

4. Experimental Results

The method was evaluated on synthetic benchmarks (Real20, Nature, SIR2 subsets) and real-world datasets (OpenRR-1K).

• Quantitative Performance:
  • ReflexSplit achieves State-of-the-Art (SOTA) results across most metrics.
  • On the OpenRR-1K real-world benchmark, it achieves the highest SSIM (0.9372) and lowest LPIPS (0.1087), indicating superior perceptual quality and structural fidelity compared to competitors like RDNet, DSIT, and DSRNet.
  • It outperforms RDNet on 63% of test images in the OpenRR-1K dataset.
• Qualitative Performance:
  • Visual comparisons show ReflexSplit effectively removes reflections while preserving fine-grained textures, sharp boundaries, and color fidelity.
  • Competing methods often exhibit residual reflections, over-smoothing, or color distortion, particularly in challenging scenarios like strong specular highlights or ambiguous scenes (e.g., a painting misidentified as a reflection).
• Ablation Studies:
  • Removing CrGF leads to catastrophic failure in feature consistency.
  • Removing LFSB (specifically the differential operator) causes severe transmission-reflection confusion (PSNR drops significantly).
  • The full Curriculum Training strategy yields better convergence than fixed-strength or single-stage scheduling.

5. Significance

ReflexSplit addresses a fundamental limitation in Single Image Reflection Separation: the progressive entanglement of features in deep networks. By introducing an explicit mechanism to alternate between sharing information (fusion) and suppressing interference (differential separation), it achieves robust generalization in complex, nonlinear mixing scenarios. The integration of curriculum learning ensures that the network learns to separate layers without sacrificing the reconstruction of shared structures, setting a new benchmark for both synthetic and real-world reflection removal tasks. The code is publicly available, facilitating further research in this domain.