PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Constrained Least Squares for Blind Image Super-Resolution

Ziwei Luo, Haibin Huang, Lei Yu, Youwei Li, Haoqiang Fan, Shuaicheng Liu

arXiv:2202.07508v3eess.IVcs.CV

TL;DR

Blind image super-resolution must recover a high-resolution image from a low-resolution input with an unknown blur kernel, while existing approaches face kernel-estimation ambiguity. The paper reformulates degradation to estimate kernels in low-resolution space and combines dynamic kernel estimation with feature-domain deblurring. It reports state-of-the-art blind SR performance across various kernels and noises.

  • Problem

    Blind SR must recover an HR image from an LR input with an unknown kernel, while estimating HR-space kernels from LR images is ambiguous after undersampling.

    Source: §1 · pp. 1–2

  • Method

    The method reformulates degradation for LR-space kernel estimation, uses a dynamic deep linear kernel, and applies DCLS to deblur features.

    Source: §1 · p. 2

  • Results

    The method achieves state-of-the-art blind SR performance across various kernels and noises.

    Source: §1 · p. 2, §5 · p. 8

  • Takeaways & Limitations

    The reformulated kernel and feature-domain deconvolution provide a principled route to clean features for blind SR reconstruction.

    Source: §1 · p. 2

Abstract

from arXiv · show

In this paper, we tackle the problem of blind image super-resolution(SR) with a reformulated degradation model and two novel modules. Following the common practices of blind SR, our method proposes to improve both the kernel estimation as well as the kernel-based high-resolution image restoration. To be more specific, we first reformulate the degradation model such that the deblurring kernel estimation can be transferred into the low-resolution space. On top of this, we introduce a dynamic deep linear filter module. Instead of learning a fixed kernel for all images, it can adaptively generate deblurring kernel weights conditional on the input and yield a more robust kernel estimation. Subsequently, a deep constrained least square filtering module is applied to generate clean features based on the reformulation and estimated kernel. The deblurred feature and the low input image feature are then fed into a dual-path structured SR network and restore the final high-resolution result. To evaluate our method, we further conduct evaluations on several benchmarks, including Gaussian8 and DIV2KRK. Our experiments demonstrate that the proposed method achieves better accuracy and visual improvements against state-of-the-art methods.

1. Introduction

The section frames blind single-image super-resolution as recovering a high-resolution image from one low-resolution observation despite an unknown blur kernel. It proposes moving kernel estimation into low-resolution space and pairing that reformulation with dynamic-kernel estimation and feature deblurring.

1. Introduction

  • Kernel estimation in high-resolution space is ambiguous after undersampling, and kernel mismatch can cause performance drops and artifacts.Kernel stretching also destroys spatial relations, making it harder for the reconstruction network to use kernel and low-resolution information together.Source: §1 · p. 2
  • The reformulated degradation model transfers blur-kernel estimation to low-resolution space, where the kernel can be estimated without aliasing ambiguity.The proposed model is described as consistent with the classical degradation model.Source: §1 · p. 2
  • The dynamic deep linear kernel provides more equivalent choices for an optimal kernel than a single-layer kernel, making it easier to learn.The method generates the kernel conditionally rather than relying on one fixed kernel for all images.Source: §1 · p. 2
  • The figure shows low-resolution images in the top row, corresponding primitive kernels in the middle row, and reformulated kernels in the bottom row.Compare each primitive kernel with its reformulated counterpart below it.Source: Figure 2 · p. 2
  • DCLS deblurs features channel-wise to obtain a clean high-resolution image without requiring kernel stretching.The reformulation and estimated kernel guide deconvolution in the feature domain.Source: §1 · p. 2
  • The method achieves state-of-the-art performance in blind SR across various degradation kernels.The introduction also identifies extensive experiments as evidence for this comparison.Source: §1 · p. 2

2. Related work

Related work covers non-blind SR methods for known degradations and blind SR methods that estimate unknown kernels. The cited approaches include kernel estimation, kernel-conditioned restoration, and iterative methods.

2. Related work

  • Bicubic-designed CNN methods can lose performance when the real degradation kernel differs from the predefined one.SRMD instead concatenates the LR image with a stretched blur kernel for restoration under different degradations.Source: §2 · p. 2
  • Blind SR generally estimates the unknown kernel and then restores the HR image using that kernel.KernelGAN estimates a kernel from a single image and applies it with a non-blind SR method such as ZSSR.Source: §2 · p. 2
  • IKC and DAN iteratively estimate kernels and restore HR images, but are described as time-consuming and computationally costly.DANv2 [31] uses a dual-path conditional block and supervises its estimator on the complete blur kernel.Source: §2 · p. 3

3. Method

The method reformulates blind SR so kernel estimation occurs in the low-resolution space, then combines a dynamic deep linear kernel estimator and feature-domain DCLS deblurring with a dual-path attention SR network. Deblurred and primitive features are used to restore the high-resolution image.

3.1. Degradation Model Reformulation

  • Reformulation derives a low-resolution kernel k_l from the high-resolution blur kernel k_h, separating kernel estimation from image upsampling.The method then estimates k_l and restores the high-resolution image.Source: §3.1 · p. 3

3.2. Dynamic Deep Linear Kernel

  • DDLK generates image-specific linear filters and collapses them into one kernel, offering equivalent parameter solutions where single-layer kernel estimation can be difficult and unstable.The filters are generated from the LR image and sequentially convolved into a single kernel.Source: §3.2 · p. 3

3.3. Deep Constrained Least Squares

  • DCLS computes deblurred features with a feature-specific operator, predicting channel-specific smooth filters and implicit Lagrange multipliers for feature-space regularization.The method notes that smooth filters and the Lagrange multiplier used in image space may be inconsistent in feature space.Source: §3.3 · p. 4

3.4. Dual-Path Attention Network

  • DPAN combines deblurred features with primitive blur features in an additive path to compensate for estimated-kernel errors, while applying channel attention on the deconvolved path.Reducing additive-path channels accelerates training and inference.Source: §3.4 · p. 4

4. Experiments

Experiments evaluate the method on isotropic and anisotropic Gaussian degradations, with comparisons on Gaussian8 benchmarks and DIV2KRK. DCLS achieves the reported strongest blind-SR performance, while ablations and real-degradation examples examine its components and scope.

4.1. Datasets and Implementation Details

  • Training uses 3,450 DIV2K and Flickr2K images; evaluation covers Gaussian8 benchmarks and anisotropic DIV2KRK, with PSNR and SSIM measured on luminance.The network uses five dual-path groups, each with ten DPABs and 64 channels.Source: §4.1 · pp. 5–6

4.2. Comparison with State-of-the-arts

  • On Gaussian8, DCLS performs best across datasets and kernel widths and remains ahead under added noise; on DIV2KRK, it outperforms the compared blind-SR methods.The DIV2KRK comparison reports PSNR and SSIM results; the Gaussian8 noise experiments span different noise levels.Source: §4.2 · pp. 6–7, Table 3 · p. 6
  • DDLK produces more accurate kernels than the compared image-specific methods, and the DIV2KRK visual results are described as sharper and cleaner.The paper also reports sharper edges on a historic real-degradation image, where HR ground truth and blur kernels are unavailable.Source: Figure 7 · p. 6, §4.2 · p. 8

4.3. Analysis and Discussions

  • The ablation reports a +0.15 dB gain from baseline when DCLS deconvolution uses the estimated kernel and high-level deep-feature information.The DPAN baseline is described as eliminating artifacts from the kernel before the DCLS improvement.Source: §4.3 · p. 8, Table 5 · p. 8
  • On real-degradation data without HR ground truth or known blur kernels, DCLS produces sharper edges and visually pleasing results than LapSRN and DANv2.The example is a historic image.Source: §4.3 · p. 8, Figure 11 · p. 8

5. Conclusion

The conclusion presents a blind-SR method that estimates a reformulated blur kernel in low-resolution space, improves kernel estimation and feature deconvolution, and reports state-of-the-art performance across tested kernels and noise levels.

5. Conclusion

  • The method derives a blur kernel in low-resolution space from the classical degradation model for blind super-resolution.It estimates and applies that kernel in HR image restoration.Source: §5 · p. 8
  • DDLK improves kernel estimation, while DCLS integrates the blur kernel and LR image in the feature domain to obtain a clean feature for the dual-path SR network.The clean feature and primitive feature are fed into the network to generate the super-resolved image.Source: §5 · p. 8
  • Experiments across various kernels and noise levels demonstrate state-of-the-art blind-SR performance.Source: §5 · p. 8

Related papers

Loading 2202.07508v3…