基于频域自适应增强稀疏先验与图正则化的盲图像去模糊方法
Blind Image Deblurring Method Based on Adaptive Frequency-Domain Enhanced Sparse Prior and Graph Regularization

盲图像去模糊旨在从模糊观测中恢复清晰图像，且不依赖任何模糊核先验知识。与依赖大规模数据训练的深度学习方法不同，该问题本质上属于小样本逆问题，其核心挑战在于如何从单幅图像或有限样本中有效估计未知的模糊核，而非通过大量训练数据学习模糊与清晰映射关系。本文基于卷积操作会降低图像高频稀疏性这一观测，提出一种面向模糊核估计的新型频域自适应增强正则化先验(Adaptive Frequency-Domain Tuning for Sparse-Based Prior, AFTS)。该先验通过频域自适应增强与非线性激活机制，有效捕捉图像模糊过程中的高频特征退化规律。我们将AFTS先验嵌入最大后验概率估计框架，构建了清晰图像与模糊核的联合优化模型，并采用半二次分裂与坐标下降策略实现高效求解。在多个标准数据集上的实验表明，本方法在PSNR、SSIM等客观指标上与主流盲去模糊算法性能相当，且在计算效率方面具备明显优势，为以后高效图像的复原提供了可行解决方案。
Blind image deblurring aims to recover sharp images from blurred observations without relying on any prior knowledge of the blur kernel. Unlike deep learning methods that depend on large-scale data training, this problem is inherently a few-shot inverse problem. Its core challenge lies in effectively estimating the unknown blur kernel from a single image or limited samples, rather than learning the mapping relationship between blurred and sharp images through massive training data. Based on the observation that convolution operations reduce the high-frequency sparsity of images, this paper proposes a novel adaptive frequency-domain enhancement regularization prior for blur kernel estimation (Adaptive Frequency-Domain Tuning for Sparse-Based Prior, AFTS). This prior effectively captures the degradation law of high-frequency features during image blurring through adaptive frequency-domain enhancement and a nonlinear activation mechanism. We embed the AFTS prior into the maximum a posteriori (MAP) estimation framework, construct a joint optimization model for sharp images and blur kernels, and adopt the half-quadratic splitting (HQS) and coordinate descent strategies to achieve efficient solutions. Experiments on multiple standard datasets demonstrate that the proposed method achieves performance comparable to state-of-the-art blind deblurring algorithms in objective metrics such as PSNR and SSIM, while exhibiting significant advantages in computational efficiency. It provides a feasible solution for high-efficiency image restoration in future applications.