TM-BSN: Triangular-Masked Blind-Spot Network for Real-World Self-Supervised Image Denoising

Junyoung Park; Youngjin Oh; Nam Ik Cho

eess.IVApr 6, 2026

TM-BSN: Triangular-Masked Blind-Spot Network for Real-World Self-Supervised Image Denoising

Authors: Junyoung Park, Youngjin Oh, Nam Ik Cho

Organizations: Department of ECE, INMC, Seoul National University, South Korea · IPAI, Seoul National University, South Korea

Abstract

Blind-spot networks (BSNs) enable self-supervised image denoising by preventing access to the target pixel, allowing clean signal estimation without ground-truth supervision. However, this approach assumes pixel-wise noise independence, which is violated in real-world sRGB images due to spatially correlated noise from the camera's image signal processing (ISP) pipeline. While several methods employ downsampling to decorrelate noise, they alter noise statistics and limit the network's ability to utilize full contextual information. In this paper, we propose the Triangular-Masked Blind-Spot Network (TM-BSN), a novel blind-spot architecture that accurately models the spatial correlation of real sRGB noise. This correlation originates from demosaicing, where each pixel is reconstructed from neighboring samples with spatially decaying weights, resulting in a diamond-shaped pattern. To align the receptive field with this geometry, we introduce a triangular-masked convolution that restricts the kernel to its upper-triangular region, creating a diamond-shaped blind spot at the original resolution. This design excludes correlated pixels while fully leveraging uncorrelated context, eliminating the need for downsampling or post-processing. Furthermore, we use knowledge distillation to transfer complementary knowledge from multiple blind-spot predictions into a lightweight U-Net, improving both accuracy and efficiency. Extensive experiments on real-world benchmarks demonstrate that our method achieves state-of-the-art performance, significantly outperforming existing self-supervised approaches. Our code is available at https://github.com/parkjun210/TM-BSN.

Figures & tables

Panel of Figure 1 — Figure 1: Visualization of Effective Receptive Fields (ERFs) of different Blind-Spot Networks (BSNs): (a) AP-BSN, (b) LG-BPN, (c) AT-BSN, and (d) the proposed TM-BSN. The ERFs are computed by backpropagating the gradients from the central output pixel with respect to the input pixels [ 24 ] .

Panel of Figure 2 — Figure 2: (a) The demosaicing filter assigns higher weights to spatially closer samples during color reconstruction. (b) Real noise exhibits a diamond-shaped correlation pattern with respect to relative distance.

Figure 3: Illustration of receptive field expansion and blind-spot formation using the proposed triangular-masked convolution. The receptive field progressively expands in a triangular shape through stacked masked convolutions, and a feature-map shift is subsequently applied to introduce a central blind-spot.

Figure 4: Overview of the proposed Triangular-Masked Blind-Spot Network (TM-BSN) architecture. R rotates the input image by $0^{\circ}$ , $90^{\circ}$ , $180^{\circ}$ , and $270^{\circ}$ and extracts features through TM-BSN; S applies feature-map shifts by a given offset $s$ to generate blind spots; and R -1 unrotates the four branches, concatenates them along the channel dimension, and applies a $1{\times}1$ convolution to produce the final output.

Methods	SIDD Validation		SIDD Benchmark		DND Benchmark
Methods	PSNR ↑ (dB)	SSIM ↑	PSNR ↑ (dB)	SSIM ↑	PSNR ↑ (dB)	SSIM ↑
Non-Learning
BM3D [ 6 ]	31.75	0.706	25.65	0.685	34.51	0.851
WNNM [ 9 ]	26.31	0.524	25.78	0.809	34.67	0.865
Supervised
DnCNN [ 43 ]	37.73	0.943	37.61	0.941	37.90	0.943

Table 1: Quantitative comparison with state-of-the-art methods on the SIDD validation, SIDD benchmark, and DND benchmark. We highlight the best and second-best results among self-supervised methods.

\dagger

denotes results from a fully self-supervised setting on the DND benchmark (right side of the slash).

\ddagger

denotes results on the SIDD benchmark evaluated via the Kaggle competition using public checkpoints, as the official website is currently unavailable.

Panel of Figure 5 — Figure 5: Qualitative comparison on SIDD Validation dataset [ 1 ] .

Panel of Figure 6 — Figure 6: Qualitative comparison on DND Benchmark dataset [ 30 ] .

Method	Params ↓	FLOPs ↓	Runtime ↓	PSNR ↑
Method	(M)	(G)	(ms)	(dB)
AP-BSN [ 20 ]	0 3.66	4456.76	287.32	35.76
PUCA [ 14 ]	12.78	2637.37	429.11	37.49
C-BSN [ 15 ]	0 3.66	481.54	34.49	36.22
SDAP (E) [ 29 ]	0 3.66	956.85	51.51	37.30
SASL [ 22 ]	0 1.08	34.93	4.60	37.39

Table 2: Comparison of model complexity and performance on the SIDD validation dataset. FLOPs and runtime are measured on

256\times 256

input patches using an NVIDIA GeForce RTX 3090 GPU.

Figure 7: Ablation study on the training shift offset $s$ . Each curve shows the PSNR on the SIDD Validation dataset as the inference shift offset is varied.

Shift Set $S$	PSNR ↑ (dB)	SSIM ↑
$\{0,1,\ldots,6\}$	37.80	0.950
$\{0,1,\ldots,7\}$	37.79	0.949
$\{1,2,\ldots,6\}$	37.81	0.951
$\{1,2,\ldots,7\}$	37.76	0.949
$\{2,3,\ldots,6\}$	38.08	0.952
$\{2,3,\ldots,7\}$	37.91	0.950

Table 3: Ablation study on the inference shift offset set

S

for knowledge distillation. We report PSNR and SSIM on the SIDD validation set. The best results are highlighted in bold .

Panel of Figure 8 — Figure 8: Qualitative comparison of denoising results with different shift offsets and the distilled TM-BSN (D).

Self-Supervised Image Denoising

Self-supervised image denoising learns to remove noise from images without requiring clean training targets.

See all 25 papers

Frames2Residual: Spatiotemporal Decoupling for Self-Supervised Video Denoising
Mar 11, 2026Mingjie Ji, Zhan Shi, Kailai Zhou +2Self-Supervised Learning Image Restoration
Nanjing University, Nanjing, China · Nanyang Technological University, Singapore · Key Laboratory of Optoelectronic Devices and Systems with Extreme Performances of MOE, Nanjing University, Nanjing, China
LoTA-N2N: Local Trace Adaptation for Zero-Shot Self-Supervised Image Denoising
Jul 27, 2026Jintong Hu, Bin Xia, Junlin Liu +2Self-Supervised Image Denoising Self-Supervised Learning
Tsinghua University · The Chinese University of Hong Kong · University of Southern California +1
Learning to Recorrupt: Noise Distribution Agnostic Self-Supervised Image Denoising
Mar 26, 2026Brayan Monroy, Jorge Bacca, Julián TachellaSelf-Supervised Image Denoising Self-Supervised Learning
Universidad Industrial de Santander, Colombia · CNRS, ENS de Lyon, France

Mar 11, 2026J/K move · Enter open · S save

Frames2Residual: Spatiotemporal Decoupling for Self-Supervised Video Denoising

Mingjie Ji, Zhan Shi, Kailai Zhou, Zixuan Fu+1

Nanjing University, Nanjing, China · Nanyang Technological University, Singapore · Key Laboratory of Optoelectronic Devices and Systems with Extreme Performances of MOE, Nanjing University, Nanjing, China

Self-Supervised Learning Image Restoration

Open paper AskPDF arXiv

Self-supervised video denoising methods typically extend image-based frameworks into the temporal dimension, yet they often struggle to integrate inter-frame temporal consistency with intra-frame spatial specificity. Existing Video Blind-Spot Networks (BSNs) require noise independence by masking the center pixel, this constraint prevents the use of spatial evidence for texture recovery, thereby severing spatiotemporal correlations and causing texture loss. To address this, we propose Frames2Residual (F2R), a spatiotemporal decoupling framework that explicitly divides self-supervised training into two distinct stages: blind temporal consistency modeling and non-blind spatial texture recovery. In Stage 1, a blind temporal estimator learns inter-frame consistency using a frame-wise blind strategy, producing a temporally consistent anchor. In Stage 2, a non-blind spatial refiner leverages this anchor to safely reintroduce the center frame and recover intra-frame high-frequency spatial residuals while preserving temporal stability. Extensive experiments demonstrate that our decoupling strategy allows F2R to outperform existing self-supervised methods on both sRGB and raw video benchmarks.

Abstract

Self-Supervised Image Denoising

Explore similar work

Frames2Residual: Spatiotemporal Decoupling for Self-Supervised Video Denoising

LoTA-N2N: Local Trace Adaptation for Zero-Shot Self-Supervised Image Denoising

Learning to Recorrupt: Noise Distribution Agnostic Self-Supervised Image Denoising