Adaptive Language-Aware Image Reflection Removal Network

Siyan Fang; Yuntao Wang; Jinpu Zhang; Ziwen Li; Yuehuan Wang

cs.CVMar 6, 2026

Adaptive Language-Aware Image Reflection Removal Network

Authors: Siyan Fang, Yuntao Wang, Jinpu Zhang, Ziwen Li, Yuehuan Wang

Organizations: Huazhong University of Science and Technology · National University of Defense Technology

Abstract

Existing image reflection removal methods struggle to handle complex reflections. Accurate language descriptions can help the model understand the image content to remove complex reflections. However, due to blurred and distorted interferences in reflected images, machine-generated language descriptions of the image content are often inaccurate, which harms the performance of language-guided reflection removal. To address this, we propose the Adaptive Language-Aware Network (ALANet) to remove reflections even with inaccurate language inputs. Specifically, ALANet integrates both filtering and optimization strategies. The filtering strategy reduces the negative effects of language while preserving its benefits, whereas the optimization strategy enhances the alignment between language and visual features. ALANet also utilizes language cues to decouple specific layer content from feature maps, improving its ability to handle complex reflections. To evaluate the model's performance under complex reflections and varying levels of language accuracy, we introduce the Complex Reflection and Language Accuracy Variance (CRLAV) dataset. Experimental results demonstrate that ALANet surpasses state-of-the-art methods for image reflection removal. The code and dataset are available at https://github.com/fashyon/ALANet.

Figures & tables

Figure 1: The impact of language-guided reflection removal with different types of language inputs. Inaccurate language inputs result in worse outcomes than having no language. The specific language inputs for each subfigure are provided in the supplementary material.

Figure 2: Overview of the proposed ALANet, which comprises various modules that use language adaptively to remove reflections. T and R represent the transmission and reflection layers, respectively.

Figure 3: Structure of the LASB. As the core of LASB, LCAM utilizes language features from different layers to facilitate the separation of those layers.

Figure 4: Structure of the LCAM. The dashed lines indicate the scenario without language input.

Figure 5: Structure of the ALCM. The ALCM enhances the consistency between language features and visual content.

Figure 6: Structure of the LSCT and its core component, LSCA. The dashed lines in LSCA indicate the scenario without language input.

Methods	Nature (20)		Real (20)		Wild (55)		Postcard (199)		Solid (200)		Average
Methods	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
BDN (ECCV’18)	18.83	0.737	18.68	0.728	22.02	0.822	20.54	0.857	22.68	0.856	20.55	0.800
ERRNet (CVPR’19)	20.39	0.766	22.28	0.796	25.14	0.873	21.53	0.877	23.55	0.880	22.58	0.838
IBCLN (CVPR’20)	23.78	0.784	21.59	0.764	24.46	0.885	22.95	0.875	24.74	0.893	23.50	0.840
LANet (ICCV’21)	23.55	0.811	22.51	0.815	26.06	0.900	24.14	0.907	24.30	0.898	24.11	0.866
YTMT (NIPS’21)	24.08	0.814	22.68	0.798	25.24	0.888	21.86	0.880	23.79	0.887	23.53	0.853

Table 1: Quantitative comparison with SOTA methods on public datasets. Bold and underline indicate top 1st and 2nd rank, respectively.

Figure 7: Qualitative comparison with SOTA methods on the CRLAV dataset (top) and the Real dataset (bottom). ALANet excels in identifying and removing reflections in complex environments. T and R represent the transmission and reflection layers, respectively.

Methods	CRLAV (600)		Param	FLOPs
Methods	PSNR	SSIM	(M)	(G)
BDN (ECCV’18)	17.46	0.686	75.16	12.70
ERRNet (CVPR’19)	18.93	0.702	18.95	116.72
IBCLN (CVPR’20)	18.81	0.701	21.61	98.16
LANet (ICCV’21)	19.28	0.709	10.93	83.81
YTMT (NIPS’21)	18.92	0.697	76.90	110.98

Table 2: Quantitative comparison with SOTA methods on the CRLAV dataset, including parameters and FLOPs (computed for a 128×128 RGB image).

Language type for training			Language type for testing			Average
Correct	Random	None	Correct	Random	None	PSNR	SSIM
✓	×	×	✓	×	×	24.74	0.868
✓	×	×	×	✓	×	24.09	0.861
✓	×	×	×	×	✓	24.02	0.856
×	✓	×	✓	×	×	24.27	0.864
×	✓	×	×	✓	×	24.11	0.860

Table 3: Ablation experiments for different types of language inputs during training and testing on public datasets.

Figure 8: Visual effects of our ALANet with different types of language inputs. The specific language inputs for each subfigure are provided in the supplementary material.

	Degree	Incorrect		Confused		Incomplete
	Degree	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
Public datasets (494)	Slightly (25%)	24.55	0.865	24.48	0.866	24.36	0.864
	Moderately (50%)	24.35	0.862	24.29	0.861	24.22	0.861
	Severely (75%)	24.08	0.859	24.01	0.859	24.03	0.857
	Entirely (100%)	22.70	0.838	24.01	0.854	24.02	0.856
CRLAV (600)	Slightly (25%)	19.43	0.713	19.65	0.714	19.52	0.718

Table 4: Ablation experiments on the effects of varying degrees of language inaccuracy.

ALCM	LSCT	LCAM		Average		Param (M)	FLOPs (G)
		Language-guided	Channel	PSNR	SSIM
		attention	attention	PSNR	SSIM
✓	✓	✓	✓	24.74	0.868	8.69	32.92
✓	✓	✓	×	24.52	0.865	8.46	32.91
✓	✓	×	✓	24.05	0.859	8.24	32.92
✓	✓	×	×	24.11	0.858	8.01	32.90

Table 5: Ablation experiments for different modules on public datasets. FLOPs for a 128x128 RGB image.

Figure 9: Feature maps in LCAM and corresponding language-guided attention weights, with the language description: “A star-shaped toy on the tabletop.”

Figure 10: Comparison of Pre-ALCM and Post-ALCM similarities between language and corresponding image features across different datasets.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 11: Multi-receptive field decoupling module.

Figure 12: Internal attention maps of the LCAM for transmission layer separation. Language description: “A building under the sky.”

Language input		Average
Transmission layer	Reflection layer	PSNR	SSIM
✓	✓	24.74	0.868
✓	×	24.60	0.866
×	✓	24.36	0.862
×	×	24.02	0.856

Appendix

Table 6: Ablation experiments on the number of language inputs.

Figure 13: Effects of our ALANet with different language numbers. T and R represent the transmission and reflection layers, respectively.

Level	Level	Level	Level	Level	Average		Param	FLOPs
0	1	2	3	4	PSNR	SSIM	(M)	(G)
64	128	128	160	160	24.74	0.868	8.69	32.92
64	128	256	512	512	24.53	0.866	55.71	53.36
64	160	160	160	160	24.35	0.866	10.36	40.32
64	128	160	160	160	24.38	0.867	9.54	34.56
64	128	128	128	128	24.27	0.865	6.95	32.42

Appendix

Table 7: Ablation experiments on channel numbers.

MSE loss	gradient loss	perceptual loss	Average
$\lambda_{1}$	$\lambda_{2}$	$\lambda_{3}$	PSNR	SSIM
1.0	2.0	0.01	24.74	0.868
1.0	1.0	0.01	24.57	0.867
1.0	2.0	0.02	24.49	0.867
1.0	2.0	0.0	24.36	0.864
1.0	0.0	0.0	24.11	0.854

Appendix

Table 8: Ablation experiments on loss functions.

Figure 14: Qualitative comparison with Zhong et al.’s method using different types of language inputs.

Figure 15: Failure case of strong nighttime indoor reflection.

CRLAV	Incorrect			Confused			Incomplete
CRLAV	R-1	R-L	ME	R-1	R-L	ME	R-1	R-L	ME
Slightly	0.854	0.854	0.849	0.854	0.853	0.848	0.915	0.915	0.858
Moderately	0.494	0.493	0.479	0.498	0.495	0.475	0.659	0.659	0.495
Severely	0.153	0.151	0.207	0.156	0.153	0.209	0.233	0.233	0.220

Appendix

Table 9: Quantitative assessment of language accuracy. Higher metric values indicate greater accuracy

Dataset	Year	Pairs	Scenes	Reflection Types	Language	Accuracy
SIR 2	2017	454	115	Not Specified	×	-
Real	2018	109	109	Not Specified	×	-
Nature	2020	220	68	Not Specified	×	-
RRW	2024	14952	150	Not Specified	×	-
CRLAV (Ours)	2025	600	163	Complex	✓	Varying

Appendix

Table 10: Comparison of real-world reflection removal datasets.

Figure 16: Qualitative comparison with SOTA methods on the Solid dataset (top) and the Wild dataset(bottom).

Figure 17: Qualitative comparison with SOTA methods on the Postcard dataset.

Figure 18: Qualitative comparison with SOTA methods on the Postcard dataset.

Figure 20: Qualitative comparison of ALANet (under descriptions with varying language accuracies) and SOTA methods on the CRLAV dataset.

Figure 21: Qualitative comparison of ALANet (under descriptions with varying language accuracies) and SOTA methods on the CRLAV dataset.

Figure 23: Examples of images in the CRLAV dataset and their annotations with varying language accuracies.

Single-Image Reflection Removal

See all 9 papers

GFRRN: Explore the Gaps in Single Image Reflection Removal
Feb 26, 2026Yu Chen, Zewei He, Xingyu Liu +2Single-Image Reflection Removal Image Restoration
Zhejiang University · Huanjiang Laboratory · The Chinese University of Hong Kong
ReLo-IRR: Reflection-Guided LoRA Framework for Image Reflection Removal
Jul 3, 2026Chaoqun Wang, Yuehuan Wei, Haoxiang Cao +1Diffusion Models for Image Restoration Single-Image Reflection Removal
School of Artificial Intelligence, South China Normal University, Guangzhou, China · University of Science and Technology of China, Hefei, China
LowAux-RDNet: Low-Pass Residual Supervision with Scene-Balanced Real-World Training for Single-Image Reflection Removal
Jul 20, 2026Jizhong LiSingle-Image Reflection Removal Image Restoration
Fudan University

Feb 26, 2026J/K move · Enter open · S save

GFRRN: Explore the Gaps in Single Image Reflection Removal

Yu Chen, Zewei He, Xingyu Liu, Zixuan Chen+1

Zhejiang University · Huanjiang Laboratory · The Chinese University of Hong Kong

Single-Image Reflection Removal Image Restoration

Open paper AskPDF arXiv

Prior dual-stream methods with the feature interaction mechanism have achieved remarkable performance in single image reflection removal (SIRR). However, they often struggle with (1) semantic understanding gap between the features of pre-trained models and those of reflection removal models, and (2) reflection label inconsistencies between synthetic and real-world training data. In this work, we first adopt the parameter efficient fine-tuning (PEFT) strategy by integrating several learnable Mona layers into the pre-trained model to align the training directions. Then, a label generator is designed to unify the reflection labels for both synthetic and real-world data. In addition, a Gaussian-based Adaptive Frequency Learning Block (G-AFLB) is proposed to adaptively learn and fuse the frequency priors, and a Dynamic Agent Attention (DAA) is employed as an alternative to window-based attention by dynamically modeling the significance levels across windows (inter-) and within an individual window (intra-). These components constitute our proposed Gap-Free Reflection Removal Network (GFRRN). Extensive experiments demonstrate the effectiveness of our GFRRN, achieving superior performance against state-of-the-art SIRR methods.

Abstract

Single-Image Reflection Removal

Explore similar work

GFRRN: Explore the Gaps in Single Image Reflection Removal

ReLo-IRR: Reflection-Guided LoRA Framework for Image Reflection Removal

LowAux-RDNet: Low-Pass Residual Supervision with Scene-Balanced Real-World Training for Single-Image Reflection Removal