convolution-free deep image prior

This work addresses the limitations of traditional Deep Image Prior (DIP)—namely its susceptibility to overfitting noise due to over-parameterization and high computational cost stemming from convolutional operations—by proposing Pool-DIP, a novel convolution-free DIP architecture. Pool-DIP uniquely integrates pooling mechanisms into contextual modeling, leveraging a pure multilayer perceptron combined with pooling operations to efficiently capture spatial information. This design substantially reduces both parameter count and computational complexity while enhancing denoising stability and generalization capability. Spectral analysis further demonstrates its effective regulation of high-frequency components. Extensive experiments show that Pool-DIP achieves competitive denoising performance on multiple synthetic and real-world datasets and successfully generalizes to super-resolution and image inpainting tasks.

To address pervasive quality degradation in ultra-high-definition (UHD) images caused by extreme resolution, this paper presents the first systematic deep learning survey dedicated to UHD image restoration. We propose a unified taxonomy that structurally categorizes restoration tasks—including super-resolution, deblurring, low-light enhancement, dehazing, deraining, and desnowing—based on network architecture and sampling strategies. We explicitly identify three UHD-specific challenges: prohibitive computational overhead, GPU memory bottlenecks, and limited generalizability under real-world degradation. Key evolutionary directions are distilled as multi-scale modeling, lightweight inference, and realistic degradation modeling. Innovatively, we establish a knowledge-graph-based survey framework and open-source a UHD-specific benchmark dataset, code repository, and curated resource list—thereby enabling standardized evaluation and practical deployment. This work has emerged as the authoritative reference in the field.

This study systematically evaluates NAFNet’s performance in image denoising and deblurring, focusing on the functional mechanisms of its core components: SimpleGate activation, Simplified Channel Attention (SCA), and LayerNorm. Through controlled ablation experiments on CIFAR-10, we quantitatively demonstrate that SimpleGate substantially outperforms conventional activations (e.g., ReLU), SCA maintains effective attention modeling while reducing parameter count, and LayerNorm significantly enhances training stability and convergence speed. Joint integration of these components yields consistent improvements—1.2–2.3 dB PSNR gain and 0.015–0.028 SSIM increase—over baseline models. To our knowledge, this is the first work to disentangle and quantify the individual contributions of NAFNet’s lightweight architectural elements. Our findings provide empirical guidance for designing efficient, stable image restoration networks, offering concrete evidence for component-level architectural decisions in practical deployment scenarios.

本文针对图像修复中的持续学习问题，提出了一种名为RwF的框架，通过为每种新退化学习一个轻量级适配器来避免遗忘，从而在成本较低的情况下有效处理多种图像退化。

GAN priors suffer from significant representation errors in inverse problems (e.g., compressed sensing, super-resolution) due to distribution mismatch between the generated and true data distributions—degrading performance on both in-distribution and out-of-distribution images. To address this, we propose a training-free hybrid modeling framework that linearly couples a fixed pre-trained GAN prior with a parameter-free, under-parameterized untrained deep decoder, jointly optimizing their combination weights in an unsupervised setting. This is the first work to synergistically integrate GAN priors and untrained decoders without task-specific GAN fine-tuning, achieving strong generalizability and scalability. Experiments on compressed sensing and super-resolution demonstrate that our method achieves substantially higher PSNR than either component used independently, while maintaining robust performance across both in-distribution and out-of-distribution test images.

This work addresses the instability of Deep Image Prior (DIP) in image reconstruction under noise overfitting when no training data are available. The authors propose a novel early-stopping framework based on self-referenced pseudo-images, which introduces the concept of dual-noise observation into DIP for the first time. By constructing a self-referenced pseudo-image from a single observation, the method enables overfitting detection without requiring accurate noise estimation. Leveraging analysis of running variance fluctuations, three early-stopping strategies tailored to inverse imaging problems are developed. Extensive experiments demonstrate that the proposed approach consistently outperforms existing DIP-based early-stopping methods across various noise types and levels, achieving superior performance in both natural image restoration and medical image reconstruction tasks.

Existing unified image restoration models suffer from high computational costs, optimization challenges due to task heterogeneity, and insufficient frequency-aware modeling. This work proposes DRNet, which introduces a novel dynamic reparameterization mechanism during initialization. By integrating a task-specific modulator-guided Dynamic Reparameterized MLP (DRMLP) with a Continuous Wavelet Transform Encoder (CWTE), DRNet enables efficient unified modeling that eliminates per-input computational redundancy, effectively fuses task-specific and general-purpose representations, and explicitly captures frequency-domain information. The method achieves state-of-the-art performance across five image restoration tasks, combining the strengths of both blind foundational models and user-guided expert models while demonstrating significantly higher parameter efficiency than current approaches.

为解决深度神经网络在资源受限环境中的高计算需求问题，提出了一种新的层合并策略，该方法能有效合并具有填充的卷积层且不增加核大小。

This study addresses the challenge that increasing resolution in imaging inverse problems often degrades the generalization performance of deep networks. The authors systematically investigate the generalization behavior of U-Net and its neural operator variants across varying discretization resolutions. Through interpretable one-dimensional models and two-dimensional limited-angle computed tomography reconstruction experiments, they find that although neural operator-based U-Nets are theoretically resolution-invariant, conventional U-Nets exhibit superior robustness and practical generalization. This work highlights a notable gap between theoretical resolution invariance and empirical performance, offering new insights for architecture selection in high-resolution inverse problem solving.

This work addresses the performance degradation of Deep Image Prior (DIP) in hyperspectral image denoising due to overfitting, which typically necessitates early stopping. To overcome this limitation, we propose a stable, unsupervised training scheme that eliminates reliance on early stopping by jointly integrating a smooth ℓ₁ data fidelity term and divergence-based sensitivity regularization into the DIP framework, complemented by input optimization. This synergistic design effectively suppresses overfitting and achieves superior denoising performance across real hyperspectral images corrupted by Gaussian, sparse, and stripe noise, outperforming existing DIP-based methods while enhancing both robustness and reconstruction quality.