Triangular-Masked Blind-Spot Network

• The paper introduces a novel triangular-masked convolution that forms a diamond-shaped blind spot to align with demosaicing-induced noise correlations in sRGB images.
• It employs a U-Net backbone that maintains full spatial resolution and uses knowledge distillation to integrate multiple blind-spot outputs into a lightweight model.
• Benchmark results on SIDD and DND datasets demonstrate state-of-the-art denoising performance with improved efficiency, higher PSNR, and effective texture preservation.

Searching arXiv for the specified paper and closely related blind-spot denoising work to ground the article in current literature. Triangular-Masked Blind-Spot Network (TM-BSN) is a blind-spot architecture for real-world self-supervised image denoising that is designed for the setting in which pixel-wise noise independence does not hold in sRGB images because of the camera image signal processing pipeline. The method is motivated by the observation that demosaicing induces spatially correlated noise with a diamond-shaped pattern, and it addresses this by replacing conventional blind-spot constructions with a triangular-masked convolution whose four rotated realizations exclude a diamond-shaped neighborhood around each pixel while retaining uncorrelated context at the original resolution. TM-BSN therefore removes the need for downsampling or post-processing, and it is coupled with a knowledge-distillation procedure that transfers multiple blind-spot predictions into a lightweight U-Net for improved accuracy and efficiency (Park et al., 6 Apr 2026).

1. Problem setting and blind-spot rationale

Blind-spot networks enable self-supervised image denoising by preventing access to the target pixel, allowing clean signal estimation without ground-truth supervision. The standard blind-spot assumption is that noise is pixel-wise independent. TM-BSN is explicitly formulated for the case in which this assumption is violated in real-world sRGB images, where spatially correlated noise arises from the camera’s image signal processing pipeline, and in particular from demosaicing (Park et al., 6 Apr 2026).

Several self-supervised methods employ downsampling to decorrelate noise, but TM-BSN is presented as an alternative because downsampling alters noise statistics and limits the network’s ability to utilize full contextual information. The method therefore keeps processing at the original resolution and instead changes the blind-spot geometry itself. This suggests a shift from decorrelation by resampling toward decorrelation by receptive-field design.

The resulting design target is not merely exclusion of the center pixel. TM-BSN is built to exclude the spatial neighborhood that is most strongly correlated with the target pixel in real sRGB data. In the paper’s discussion, this is framed as avoiding both the “identity mapping” failure of small blind spots and the excessive information loss of large rectangular masks (Park et al., 6 Apr 2026).

2. Triangular-masked convolution

The central operator in TM-BSN is the Triangular-Masked Convolution (TMC). For a $k\times k$ convolutional kernel, the method defines a binary mask matrix

$$M \;=\; [M_{uv}]_{u,v=0}^{k-1} \quad\text{where}\quad M_{uv} \;=\; \begin{cases} 1, & u \le v,\ 0, & u > v. \end{cases}$$

This mask retains only the upper-triangular region, including the main diagonal, and zeros out the strictly lower-triangular part (Park et al., 6 Apr 2026).

For $k=3$, the mask is

$$M_{3\times3} = \begin{bmatrix} 1 & 1 & 1\ 0 & 1 & 1\ 0 & 0 & 1 \end{bmatrix}.$$

If the learnable kernel is $W=[W_{uv}]$ with scalar bias $b$, and $X$ and $Y$ denote the input and output feature maps, then TMC computes

$$Y_{i,j} \;=\; \sum_{u=0}^{k-1} \sum_{v=0}^{k-1} M_{u,v}\;W_{u,v}\;X_{i+u - \lfloor k/2\rfloor,\;j+v - \lfloor k/2\rfloor} \;+\; b,$$

or equivalently

$Y \;=\; (W\odot M)\;*\;X \;+\; b.$

Because $$M \;=\; [M_{uv}]_{u,v=0}^{k-1} \quad\text{where}\quad M_{uv} \;=\; \begin{cases} 1, & u \le v,\ 0, & u > v. \end{cases}$$0 removes all weights below the main diagonal, stacking multiple TMC layers causes the receptive field to expand only in a triangular region. The paper describes this as an upper-right growth pattern for one orientation. By combining triangular masks in all four cardinal orientations through $$M \;=\; [M_{uv}]_{u,v=0}^{k-1} \quad\text{where}\quad M_{uv} \;=\; \begin{cases} 1, & u \le v,\ 0, & u > v. \end{cases}$$1 image rotations, TM-BSN constructs a blind spot that is diamond-shaped when mapped back to the original image orientation (Park et al., 6 Apr 2026).

This operator is not a generic sparsification of convolution. Its purpose is geometric: the mask is chosen so that repeated composition of masked layers yields a blind region with a shape aligned to real sRGB noise correlation.

3. Diamond-shaped blind spot and correlation geometry

TM-BSN’s geometric claim is that the relevant correlation structure in real sRGB denoising is diamond-shaped rather than rectangular. The paper attributes this to demosaicing, in which missing color channels are interpolated from neighboring samples with weights that decay roughly linearly with distance. Empirical correlation maps, including examples from the SIDD dataset, are reported to show a diamond-shaped autocorrelation in which pixels at Manhattan distance $$M \;=\; [M_{uv}]_{u,v=0}^{k-1} \quad\text{where}\quad M_{uv} \;=\; \begin{cases} 1, & u \le v,\ 0, & u > v. \end{cases}$$2 from the center are more strongly correlated than those with greater distance (Park et al., 6 Apr 2026).

The blind-spot construction is therefore defined in terms of four triangular exclusion wedges. A single upper-triangular mask at orientation $$M \;=\; [M_{uv}]_{u,v=0}^{k-1} \quad\text{where}\quad M_{uv} \;=\; \begin{cases} 1, & u \le v,\ 0, & u > v. \end{cases}$$3 excludes all pixels strictly below the main diagonal of the local patch. Rotating the image by $$M \;=\; [M_{uv}]_{u,v=0}^{k-1} \quad\text{where}\quad M_{uv} \;=\; \begin{cases} 1, & u \le v,\ 0, & u > v. \end{cases}$$4 and applying the same mask excludes a different triangular region, originally the top-left corner. Repeating this for $$M \;=\; [M_{uv}]_{u,v=0}^{k-1} \quad\text{where}\quad M_{uv} \;=\; \begin{cases} 1, & u \le v,\ 0, & u > v. \end{cases}$$5 and $$M \;=\; [M_{uv}]_{u,v=0}^{k-1} \quad\text{where}\quad M_{uv} \;=\; \begin{cases} 1, & u \le v,\ 0, & u > v. \end{cases}$$6 yields four triangular exclusions which together carve out a central diamond. A final feature-shift operation by $$M \;=\; [M_{uv}]_{u,v=0}^{k-1} \quad\text{where}\quad M_{uv} \;=\; \begin{cases} 1, & u \le v,\ 0, & u > v. \end{cases}$$7 pixels in vertical and horizontal directions expands that diamond, producing a blind-spot diamond of radius $$M \;=\; [M_{uv}]_{u,v=0}^{k-1} \quad\text{where}\quad M_{uv} \;=\; \begin{cases} 1, & u \le v,\ 0, & u > v. \end{cases}$$8 (Park et al., 6 Apr 2026).

The significance of this construction is that the excluded set is intended to coincide with the high-correlation neighborhood created by demosaicing, while the retained set preserves uncorrelated context. In the paper’s formulation, TM-BSN “precisely excludes those high-correlation pixels while retaining the remaining context.” A plausible implication is that the method’s performance depends not only on blind-spot size but on geometric agreement between the blind spot and the noise autocorrelation.

4. U-Net realization at original resolution

TM-BSN uses a U-Net-style backbone, specifically the VDIR backbone, “to avoid pooling misalignment,” and every standard $$M \;=\; [M_{uv}]_{u,v=0}^{k-1} \quad\text{where}\quad M_{uv} \;=\; \begin{cases} 1, & u \le v,\ 0, & u > v. \end{cases}$$9 convolution is replaced by a Triangular-Masked Convolution (Park et al., 6 Apr 2026). The architecture keeps full spatial resolution throughout.

A typical layer configuration is described as follows. In the encoder stage, TM-BSN uses a TMC with 64 filters and ReLU, followed by another TMC with 64 filters and ReLU, and downsampling is not performed. In the bottleneck, several TMC layers are used with 128 filters each. In the decoder stage, the network again uses TMC layers with 64 filters and ReLU. Instead of conventional upsampling, the final feature map is shifted up by $k=3$0 pixels and right by $k=3$1 pixels, with zero-padding and cropping to preserve spatial size (Park et al., 6 Apr 2026).

The four rotated branches share the TMC-based feature extractor. Only the final feature-shift and a $k=3$2 fusion convolution, used to merge the four branches, differ. This design means that no explicit downsampling or upsampling by pooling, unpooling, or pixel-shuffle is required.

The architectural consequences are stated directly in the paper’s discussion: “No downsampling/pixel-shuffle means no checkerboard artifacts, no domain gap between training/inference, and full preservation of high-frequency texture” (Park et al., 6 Apr 2026). In relation to earlier blind-spot practice, this positions TM-BSN as a same-resolution alternative to decorrelation-by-resampling.

5. Knowledge distillation and computational profile

TM-BSN can produce multiple blind-spot outputs by varying the shift offset $k=3$3, at about $k=3$4 compute cost. These outputs are treated as teacher predictions and distilled into a small non-blind U-Net with 1.02 M parameters, described as identical to AT-BSN’s student (Park et al., 6 Apr 2026).

The teacher outputs are

$k=3$5

and the student produces a prediction $k=3$6. The Recharged Distillation loss is

$k=3$7

where $k=3$8 is a random binary mask and $k=3$9 denotes stop-gradient (Park et al., 6 Apr 2026).

The student is trained for 200 k iterations on $$M_{3\times3} = \begin{bmatrix} 1 & 1 & 1\ 0 & 1 & 1\ 0 & 0 & 1 \end{bmatrix}.$$0 patches with batch size 8, using Adam with initial learning rate $$M_{3\times3} = \begin{bmatrix} 1 & 1 & 1\ 0 & 1 & 1\ 0 & 0 & 1 \end{bmatrix}.$$1 for 100 k iterations and cosine annealing to $$M_{3\times3} = \begin{bmatrix} 1 & 1 & 1\ 0 & 1 & 1\ 0 & 0 & 1 \end{bmatrix}.$$2 for the remainder (Park et al., 6 Apr 2026). The distillation objective is not only model compression. The paper states that it transfers “complementary knowledge from multiple blind-spot predictions” into the lightweight U-Net, improving both accuracy and efficiency.

On an input of size $$M_{3\times3} = \begin{bmatrix} 1 & 1 & 1\ 0 & 1 & 1\ 0 & 0 & 1 \end{bmatrix}.$$3 on RTX3090, the reported efficiency figures are: AP-BSN with 3.66 M parameters, 4457 G FLOPs, 287 ms, and 35.76 dB; AT-BSN (D) with 1.27 M, 164.6 G, 14.8 ms, and 36.80 dB; TM-BSN with 1.35 M, 633.7 G, 61.4 ms, and 37.31 dB; and TM-BSN (D) with 1.02 M, 26.7 G, 3.21 ms, and 38.08 dB (Park et al., 6 Apr 2026). These figures indicate that the distilled model is both smaller and faster than the teacher while also improving PSNR on the cited SIDD validation setting.

6. Benchmarks, relation to earlier blind-spot designs, and stated limitations

The experiments use SIDD Medium, with 320 training images and evaluation on SIDD-val with 1 280 patches of size $$M_{3\times3} = \begin{bmatrix} 1 & 1 & 1\ 0 & 1 & 1\ 0 & 0 & 1 \end{bmatrix}.$$4, and the SIDD benchmark; and DND with 50 scenes, evaluated both in the regime trained on SIDD and tested on DND and in the fully self-supervised regime on DND (Park et al., 6 Apr 2026).

The reported self-supervised denoising results are summarized below.

Setting	Method	Reported result
SIDD validation	APR (RD)	38.00 dB PSNR, 0.947 SSIM
SIDD validation	TM-BSN (teacher only, $$M_{3\times3} = \begin{bmatrix} 1 & 1 & 1\ 0 & 1 & 1\ 0 & 0 & 1 \end{bmatrix}.$$5)	37.31 dB, 0.940 SSIM
SIDD validation	TM-BSN (D, distilled)	38.08 dB, 0.952 SSIM
SIDD benchmark	APR (RD)	38.26 dB, 0.895 SSIM
SIDD benchmark	TM-BSN (D)	38.31 dB, 0.900 SSIM
DND benchmark	TBSN	39.08 dB, 0.945 SSIM
DND benchmark	TM-BSN (D)	39.41 dB, 0.949 SSIM

The paper characterizes these outcomes as state-of-the-art performance on real-world benchmarks for self-supervised denoising (Park et al., 6 Apr 2026). The comparison with TBSN is particularly relevant because TBSN is another blind-spot architecture for real-world self-supervised denoising, but it redesigns channel and spatial attentions to satisfy the blind-spot constraint in a Transformer-based network. TBSN uses grouped channel self-attention, masked window-based self-attention, patch-unshuffle and patch-shuffle, and a knowledge-distillation strategy into a plain U-Net student (Li et al., 2024). On the DND benchmark, the reported TBSN result is 39.08 dB and 0.945 SSIM, whereas TM-BSN (D) is reported at 39.41 dB and 0.949 SSIM (Park et al., 6 Apr 2026).

The stated limitations and future directions are explicit. The current mask is “strictly linear (upper triangular) but real ISP filters may have non-uniform weights”; one possible extension is to learn a smooth mask shape. Additional directions are extending the approach to CFA-domain denoising by designing masks that mirror Bayer-pattern anisotropy, and combining the method with attention or transformer blocks to enlarge unmasked context adaptively (Park et al., 6 Apr 2026).

Taken together, TM-BSN occupies a specific position within blind-spot denoising research: it preserves the self-supervised blind-spot principle, but changes the blind-spot geometry to match the diamond-shaped correlation pattern induced by demosaicing. In that sense, its primary contribution is not blind-spot denoising in general, but blind-spot denoising whose receptive-field exclusion is explicitly aligned with real sRGB noise structure.