Blur Pattern Pretraining (BPP)

• Blur Pattern Pretraining (BPP) is a framework that models diverse blur patterns using generative techniques like a VAE-based kernel generator to improve blind deblurring.
• It integrates progressive blur curricula and simulation-based priors, achieving up to 5 dB PSNR improvements over conventional deblurring methods.
• BPP leverages large-scale synthetic simulations and multi-modal guidance (motion and semantic), enhancing robustness to real-world blur and enabling accurate image recovery.

Blur Pattern Pretraining (BPP) is a methodological framework for learning, representing, and exploiting diverse blur patterns in both image restoration and robust representation learning. BPP spans generative kernel modeling for blind deblurring, large-scale simulation-based blur prior learning for generalizable deblurring, and progressive blur-based curricula for robust image classification. Its core rationale is that explicit modeling and staged exposure to blur diversity engender models that are resilient to real-world blur and input corruptions, outperforming static augmentation and dataset-specific solutions.

1. Generative Modeling of Blur Kernels: VAE-Based BPP

The blur-kernel generative prior approach (Asim et al., 2019) encodes the space of motion blur kernels using a Variational AutoEncoder (VAE). The architecture comprises:

• Encoder: Two convolutional layers (20 filters, 2×2, stride 1, ReLU, max-pooling), a flatten step, and two parallel fully connected layers (outputting $\mu(z|x)$ and $\log\,\sigma(z|x)$) with latent dimensionality $m=50$.
• Decoder (kernel generator $G_K$): Accepts latent $z_k\in\mathbb{R}^{50}$, expands to 720 via FC, reshapes to feature maps, followed by two upsampling and transposed-convolution stages (20 filters, 2×2, ReLU), and a final transposed-convolution (1 filter, ReLU) to yield $k(z_k)\in\mathbb{R}^{25\times25}$.

Pretraining employs 80,000 synthetic blur kernels comprising straight and curved motion blurs (length 5–28 px), normalized to sum to 1 and fixed spatial support (zero-padding as required). The VAE is trained with the standard ELBO:

$$L(\phi, \theta) = \mathbb{E}_{z\sim q_\phi(z|x)}\big[\|x-G_K(z)\|_2^2\big] + \mathrm{KL}\big(q_\phi(z|x)\;\|\;\mathcal{N}(0, I)\big)$$

Optimized via Adam (lr $1\times10^{-5}$, batch size 5, $\sim$100 epochs).

Sampling $z_k\sim\mathcal{N}(0,I)$ yields normalized kernels spanning support, orientation, and length variation. Empirical observations: latent norm $\|z_k\|$ correlates with blur extent (larger norm, longer PSF).

2. Integration of Pretrained Blur Priors in Blind Deblurring

BPP’s VAE-based blur prior is deployed as a generative regularizer in blind image deconvolution. The latent-space optimization objective is:

$$\min_{z_i\in\mathbb{R}^l,\,z_k\in\mathbb{R}^m} \|y - G_I(z_i)\otimes G_K(z_k)\|_2^2 + \gamma\|z_i\|_2^2 + \lambda\|z_k\|_2^2$$

with $G_I$ denoting a pretrained image generator (GAN/VAE), $G_K$ the blur generator, and $\otimes$ denoting convolution.

Optimization proceeds via alternating gradient descent, with exponential step-size decay and optional random restart. This process estimates both deblurred image and blur kernel, with optional slack/prior relaxation enabled using an auxiliary image variable and TV regularization.

Empirically, BPP yields $\sim$2–5 dB PSNR and 0.1 SSIM gain over classic priors and end-to-end CNN baselines, and maintains performance as blur length/noise increases. Range-error ablation demonstrates BPP-recovered kernels approach the VAE expressive upper bound, with PSNR gap $\leq$2 dB. Visual comparisons confirm accurate recovery of blur kernel orientation, length, and structure.

3. Progressive Blur Curriculum: Human-Inspired BPP for Robust Representation

BPP as implemented in the Visual Acuity Curriculum (VAC) (Raj et al., 16 Dec 2025) structurally mimics human infant vision development. Training commences on highly blurred image inputs, progressively reducing the blur:

• Blur Operation: $x̃ = G_\sigma * x$, with $G_\sigma$ as a 2D Gaussian kernel ($\sigma$ determining blur radius).
• Blur Schedule: Defined by

$$\sigma(t) = \begin{cases} \sigma_k & \text{if } \sum_{i=0}^{k-1} n_i < t \leq \sum_{i=0}^k n_i \ 0 & \text{if } t > \sum_{i=0}^K n_i \end{cases}$$

where segments $(n_k, \sigma_k)$ are chosen for an initial “deficit” phase ($N_\text{def}=\lfloor N/5\rfloor$), followed by halving $\sigma$ per segment.

• Replay Mechanism: Throughout training, examples from previously seen blur levels are sampled probabilistically for replay, enforcing retention of low-frequency feature representations.

VAC shows a reduction of mean corruption error (mCE) by up to 8.30 pp on CIFAR-10-C and 4.43 pp on ImageNet-100-C vs vanilla training, with moderate clean-error increases. In controlled ablations, VAC outperforms static random-blur augmentation (VAC 17.58 mCE vs. constant-blur 18.02 mCE), confirming the importance of curriculum over randomization.

VAC is compatible and synergistic with MixUp, CutMix, ℓ₂-adversarial training, RandAugment, and AutoAugment, further lowering mCE and adversarial attack success rates despite the accuracy–robustness trade-off.

4. Large-Scale Simulation-Based BPP for Real-World Deblurring Generalization

Recent advances (Gao et al., 10 Jan 2026) reveal that dataset-specific training fails to generalize due to insufficient blur pattern diversity. BPP directly addresses this by:

• Stage 1: Pretraining on large synthetic simulation datasets (GSBlur: motion via random 3D camera trajectories, LSDIR: Gaussian/sample-trajectory motion kernels added to real images), encompassing broad blur pattern support—lengths (5–75 px), orientations ($0$–$360^\circ$), spatial non-uniformity, and semantic diversity.
• Loss Formulation:
  • Pretraining objective: $$\mathcal{L}_{\text{pre}}=L_{\text{rec}}+\lambda_{\text{align}}L_{\text{align}}$$ (with $L_{\text{rec}}$ pixel-wise and $L_{\text{align}}$ domain alignment).
  • Fine-tuning: $$\mathcal{L}_{\text{fine}}=\alpha L_{\text{pixel}}+\beta L_{\text{perc}}+\gamma L_{\text{adv}}+\delta \mathcal{L}_{\text{pre}}$$ (pixel, perceptual, adversarial, and pretraining regularizer terms).
  • Domain alignment penalizes centroid drift between simulated and real-data latent features.

Training uses UNet pre-reconstruction, deep-compression autoencoder for tokenization, and Linear-DiT (Transformer) blocks with O(N) attention:

$$O_i = \mathrm{ReLU}(Q_i) \frac{\sum_j \mathrm{ReLU}(K_j)^\top V_j}{\mathrm{ReLU}(Q_i) (\sum_j \mathrm{ReLU}(K_j)^\top)}$$

Enables scalable application to high-resolution inputs.

5. Enhancement via Motion and Semantic Guidance (MoSeG)

To further strengthen BPP in extremely degraded scenarios, the pipeline integrates MoSeG:

• Motion Guidance (MoG): UNet predicts dense motion offsets $\Delta P \in \mathbb{R}^{H\times W\times 2}$ from blurred input, concatenated in UNet feature maps for improved spatial activation. No additional supervision required.
• Semantic Guidance (SeG): Vision-LLM (Qwen2.5-VL) produces high-level captions $c$ for each blurred input, embedded into vector $t$, which conditions latent transformer key–value projections.

These guidance signals enable the network to exploit both pixel-level motion cues and global semantic context, anchoring restoration even where low-level cues are insufficient.

6. Evaluation and Impact Across Applications

Quantitative results demonstrate state-of-the-art improvements in cross-dataset generalization, PSNR/SSIM, and no-reference perceptual metrics (MANIQA, LIQE, NRQM, CLIP-IQA, PI, NIQE, ILNIQE) on both simulated and real datasets. GLOWDeblur (Gao et al., 10 Jan 2026), built on BPP, achieves an average PSNR gain of $\sim$1.5 dB over competing methods, with top ranks in perceptual metrics and recovery of fine details in complex scenes—such as legible text and structure in real-world images.

VAC (Raj et al., 16 Dec 2025) decreases mCE versus vanilla CNN training and static blur augmentation, and is complementary to modern augmentation and adversarial schemes.

The VAE-kernel prior (Asim et al., 2019) yields notable gains in blind deblurring robustness to blur size and noise, outperforming classical priors and end-to-end models, approaching the expressive ceiling set by the generative kernel family.

7. Significance and Conceptual Implications

BPP reframes blur modeling from dataset-centric memorization to pattern-centric abstraction. By leveraging generative models, large-scale simulation priors, staged curricula, and auxiliary signals, BPP endows models with resilience to real-world blur variability and input corruptions. This suggests BPP is instrumental both in restoration (deblurring) and robust recognition, and that maximizing pattern diversity (rather than dataset realism alone) is the principal axis for generalization.

A plausible implication is that future robust vision frameworks will increasingly exploit synthetic pattern diversity amplification, curriculum schedules inspired by biological development, and joint-guidance strategies for unstructured degradation. BPP’s principled mechanisms are extendable to further modalities where nuisance pattern diversity is critical to out-of-domain robustness.