Speckle2Void: Deep Self-Supervised SAR Despeckling with Blind-Spot Convolutional Neural Networks

I. INTRODUCTION

• SAR speckle is signal-dependent granular noise that makes images difficult to interpret and can hinder segmentation, detection, and recognition.The introduction places despeckling among spatial- and transform-domain methods, including SAR-BM3D.Source: §I · p. 1
• Clean SAR images are unavailable, so supervised CNN methods use synthetically speckled optical images or approximate multitemporal targets, each with practical drawbacks.Synthetic training can create a domain gap and artifacts or poor radiometric-feature preservation; multitemporal acquisition, registration, and temporal variation can complicate speckle rejection.Source: §I · p. 1
• Blind-spot networks reconstruct each clean pixel from neighbors while hiding the target pixel during training to prevent identity copying.Self-supervised denoising methods, use this approach with a single noisy version of each image.Source: §I · p. 1
• Speckle2Void is a self-supervised Bayesian despeckling framework that enables direct training on real SAR images.It avoids training on synthetically speckled optical images and the difficulty of constructing multitemporal datasets.Source: §I · p. 1
• The method formulates a Bayesian model for speckle and for clean-pixel priors conditioned on neighboring pixels.It also improves the blind-spot CNN architecture and uses regularized training with a variable blind-spot shape to account for speckle autocorrelation.Source: §I · pp. 1–2
• Speckle2Void outperforms model-based techniques, approaches supervised methods on synthetic data, and is superior to them on real SAR data.The paper also reports that improvements to network modeling, loss, and training resolve residual granularity seen in its preliminary version.Source: §I · p. 2

A. SAR Despeckling

• Spatial- and wavelet-domain methods use filtering, multiresolution analysis, or MAP estimation, but can introduce artifacts and preserve detail poorly.Homomorphic wavelet approaches may leave bias from the log-speckle mean in reconstructed images.Source: A. SAR Despeckling · p. 2
• Non-local methods exploit patch similarity; PPB adapts this approach to multiplicative non-Gaussian speckle, and SAR-BM3D combines non-local grouping with SAR-adapted filtering.SAR-BM3D uses PPB similarity and an LMMSE estimator in place of BM3D's hard-thresholding and Wiener filtering.Source: A. SAR Despeckling · p. 2
• Most described CNN despeckling methods use supervised training with synthetic data or multitemporal SAR targets.SAR-CNN uses averaged multitemporal images; ID-CNN and subsequent variants use synthetically speckled images.Source: A. SAR Despeckling · pp. 2–3

B. Self-supervised denoising with CNNs

• Noise2Noise needs image pairs with independent noise, whereas Noise2Void and Noise2Self use blind spots to learn from one noisy image.The blind spot excludes the center pixel from the receptive field, preventing identity mapping under the stated uncorrelated-noise assumption.Source: B. Self-supervised denoising with CNNs · p. 3
• Laine et al. process the whole image with a blind-spot CNN and introduce a Bayesian framework for noise models and conditional priors.The paper builds on blind-spot denoising with a self-supervised Bayesian framework for SAR despeckling.Source: B. Self-supervised denoising with CNNs · p. 3, A. SAR Despeckling · p. 3

III. BACKGROUND

• The blind-spot CNN predicts a clean pixel from neighboring noisy pixels while keeping that pixel hidden during training.In the Bayesian view, the CNN represents p(xi|yi, Ωyi), and the target yi is used as a noisy training target rather than to compute the clean estimate.Source: §III · p. 3
• The CNN predicts the receptive-field-conditioned pixel prior; the posterior mean is the MMSE estimate, and training maximizes the noisy-data likelihood under pixel-wise i.i.d. noise.A handcrafted prior can remain flexible because its parameters are predicted pixel by pixel by the CNN.Source: §III · p. 3

A. Model

• The model assumes multiplicative SAR observations yi = nixi, with Gamma-distributed speckle of unit mean and variance 1/L for an L-look image.Despeckling aims to estimate intensity backscatter x from observed intensity return y.Source: A. Model · p. 3

IV. PROPOSED METHOD

• Four rotated image branches compute directional receptive fields, which are shifted, rotated back, concatenated, and fused to predict inverse-Gamma parameters for each pixel.The fusion uses 1×1 convolutions.Source: Figure 1 · p. 4

A. Model

• During training, the blind-spot network estimates αxi and βxi by minimizing negative log noisy-data likelihood; at testing, the MMSE estimator combines the prior, noisy pixel, and noise parameters.The figure outlines the training and testing phases.Source: Figure 2 · p. 4
• The model uses a receptive-field-conditioned inverse-Gamma prior with CNN-predicted αxi and βxi; the posterior remains inverse-Gamma and the noisy-data likelihood is available in closed form.The noise is modeled as Gamma-distributed.Source: A. Model · p. 4

B. Training

• Whitening is used to decorrelate speckle, but perfect decorrelation is impossible and residual correlation may limit the blind-spot CNN.Whitening is especially important because neural networks can overfit random patterns.Source: B. Training · pp. 4–5
• A larger blind spot can cover the noise-autocorrelation peak, but reduces usable neighboring information and trades noise suppression for blurrier predictions.The trade-off matters because immediate neighbors are the most similar to the target pixel.Source: B. Training · p. 5
• Training alternates a 1×1 blind spot with a larger, potentially irregular one, then testing uses a 1×1 blind spot to recover high-frequency content while suppressing most noise correlation.The larger shape can be chosen to match the noise autocorrelation; nearby pixels are weakly trained so the network does not rely too heavily on them.Source: B. Training · p. 5

C. Testing

• The testing-stage MMSE estimate combines the CNN-predicted prior parameters with the noisy pixel observation.The estimate is the expected value of the posterior distribution.Source: C. Testing · p. 5

D. Loss function

• The final loss combines the negative log likelihood of noisy observations with anisotropic total variation on the posterior image.The TV weight tunes the desired degree of smoothness.Source: D. Loss function · p. 5

E. Blind-spot architecture

• Four directional receptive fields exclude the center pixel through shift-and-padding operations, with a final shift removing the center row or column.Fig. 3 depicts exclusion of the center pixel and two pixels in the vertical direction.Source: E. Blind-spot architecture · pp. 5–6, Figure 3 · p. 5
• Speckle2Void shares weights between azimuth-oriented branches and between range-oriented branches, and allows different blind-spot shifts for the two directions.This design accounts for potentially different directional statistics, including residual speckle autocorrelation.Source: E. Blind-spot architecture · p. 6

F. Non local convolutional layer and its adaptation to blind-spot networks

• Non-local layers exploit distant self-similarity to recover information hidden by the blind spot, but incur a significant computational cost.The local version uses standard convolutional layers; the non-local version adds non-local layers.Source: F. Non local convolutional layer and… · p. 6
• A mask restricts non-local comparisons to feature vectors within each directional receptive field, while blind-spot shaping occurs after the four fields are computed.The adaptation adds masking to the NLRN non-local module.Source: F. Non local convolutional layer and… · p. 6

V. EXPERIMENTAL RESULTS AND DISCUSSIONS

• The evaluation compares methods on synthetic data with ground-truth metrics and on real SAR data using no-reference metrics and visual results, then tests design choices by ablation.Source: §V · p. 6

A. Quality assessment criteria

• Synthetic-image performance is assessed with PSNR; real-image assessment uses ENL, ratio-image moments, quality index M, and ratio image structuredness (RIS).M combines deviation from ideal ENL and speckle mean with residual geometrical content; RIS measures geometrical content alone.Source: A. Quality assessment criteria · pp. 6–7

B. Reference methods

• The reference set comprises PPB, SAR-BM3D, a CNN baseline using the improved loss, and ID-CNN.Both CNN references use supervised training with synthetically speckled natural images.Source: B. Reference methods · p. 7
• SAR-CNN and are not directly compared because they use multitemporal data and their dataset is unavailable.The authors consider that setting unfair against a single observation of a scene.Source: B. Reference methods · p. 7
• The synthetic-image model uses a 1 × 1 blind spot for pixel-wise i.i.d. injected speckle, whereas the real-SAR blind spot varies during training.The architecture has 17 convolutional blocks per branch, followed by three 1 × 1 convolutions; the non-local version adds five non-local layers.Source: B. Reference methods · p. 7

C. Synthetic dataset

• The synthetic SAR-like dataset uses natural images paired with noisy versions corrupted by simulated i.i.d. speckle for a single-look image (L = 1).Source: C. Synthetic dataset · pp. 7–8
• Speckle2Void outperforms PPB and SAR-BM3D and achieves results comparable to supervised ID-CNN and CNN-baseline methods on synthetic images.The authors report similar behavior in visual assessment and a slight improvement when using non-locality.Source: C. Synthetic dataset · p. 8
• Training patches come from 450 Berkeley Segmentation Dataset images, and the method reaches comparable supervised-CNN results without clean training data.The reported training used about 400 epochs with batch size 16 and learning rate 10^-5.Source: C. Synthetic dataset · p. 8

• Real TerraSAR-X training and testing images are whitened with the blind speckle decorrelator, applied consistently across methods.Source: D. TerraSAR-X dataset · p. 8
• PPB has the best σr result but poor detail preservation, while SAR-BM3D offers higher fidelity and finer detail yet retains streaks and unrealistic texture.Source: D. TerraSAR-X dataset · pp. 9–10
• Speckle2Void produces more realistic details and higher-quality images than the reference methods, with minimal visible patterns in ratio images.The comparison includes SAR-BM3D, PPB, CNN baseline, ID-CNN, and Speckle2Void+NL.Source: Figure 5 · p. 9, D. TerraSAR-X dataset · p. 10, Figure 8 · p. 10
• Adding non-local layers marginally improves detail preservation, with lower M and RIS values, but increases computational overhead and training time.Source: D. TerraSAR-X dataset · p. 10

1) Original vs whitened:

• Whitening markedly improves the cleaned image compared with training on original TerraSAR-X data; the full method combines whitening with a variable blind spot.Fig. 9 reports ENL values of 1.28 for the original-data result, 14.5 for the whitened-data result, and 88.5 for Speckle2Void.Source: 1) Original vs whitened · p. 11, Figure 9 · p. 11

2) Enlarging the blind-spot:

• A 1×1 blind spot preserves sharper edges and more detail but leaves more residual noise; a 3×3 blind spot suppresses speckle more effectively but smooths details and edges.Speckle2Void combines stronger suppression with detail preservation; the 3×3 comparison uses that shape at test time, unlike the regularized procedure.Source: 2) Enlarging the blind-spot · p. 11, Figure 10 · p. 11

3) Effect of the TV regularizer:

• TV regularization reduces artifacts and produces smoother flat areas than training without TV.Source: 3) Effect of the TV regularizer · p. 11, Figure 11 · p. 12

4) Prior vs posterior:

• The posterior mean is visually preferable to the prior image, whose edges appear fuzzy and whose texture remains granular.The posterior estimate uses the noisy SAR observations.Source: 4) Prior vs posterior · p. 11

VI. CONCLUSION

• Speckle2Void is a self-supervised despeckling framework based on blind-spot CNNs and real single-look complex SAR images, producing faithful details without visible residual speckle.Source: §VI · pp. 11–12