Ping An

Light field imaging suffers from an inherent spatial-angular resolution trade-off. While deep learning has advanced light field image super-resolution (LFISR), convolutional neural networks (CNNs) are limited by local receptive fields that fail to capture long-range angular dependencies, and Vision Transformers introduce quadratic complexity and high-frequency detail loss. We propose FMDL-Net, a disentangled network that integrates Fourier Modulated Attention (FMA) and Dynamic Mixing Layers (DML) to address these challenges. FMA leverages the Fast Fourier Transform to achieve efficient global receptive fields and maintain geometric consistency, while DML employs content-adaptive dynamic convolutions to preserve complex textures and handle varying disparities. Experiments on five public datasets demonstrate that FMDL-Net outperforms state-of-the-art methods in PSNR and SSIM, with fewer parameters and lower computational cost.

Noise significantly degrades light field (LF) imaging quality, particularly in handheld LF cameras due to the light-scattering effects of microlens arrays. However, current LF denoising methods suffer from several limitations: (1) predominant focus on the pixel domain restricts sufficient exploration of LF frequency characteristics; (2) excessive dependence on local information or straightforward addition/concatenation for integrating global and local features impairs global structural consistency and introduces artifacts. Therefore, we propose a dual domain multi-scale network (DDMSN) that integrates frequency-domain noise priors with pixel-domain texture and epipolar geometric structure awareness. Specifically, in the frequency domain module, we propose a dual frequency collaborative enhancement (DFCE) strategy for differentiated treatment of high frequency noise characteristics and valid structural information, and adaptively enhance low frequency and high frequency band components to achieve a fine-grained restoration of the frequency domain structure. In the pixel domain module, to achieve texture detail restoration and structural consistency preservation, we introduce a global-local feature interaction (GLFI) mechanism embedded within the encoder-decoder architecture, which fuses spatial-angular local features with EPI global features through channel attention weighting, optimizing local receptive fields and global structural consistency dependencies jointly. Extensive experiments on synthetic and real-world LF noise datasets demonstrate that our method achieves significant improvements over existing methods in both qualitative visual results and quantitative metrics.

Light field (LF) denoising is pivotal for enhancing downstream LF applications. However, most existing methods focus primarily on removing noise signals, lacking sufficient preservation and utilization of the intrinsic structural consistency of LF, which leads to suboptimal denoising performance and compromised geometric relationships. To this end, we propose a collaborative multi-stage attention mechanism that integrates complementary cues from residuals, geometric constraints, and global information. Specifically, first, we design a residual-driven channel attention module to leverage residual analysis for coarse denoising; second, an epipolar-guided cross attention module to explicitly preserve structural consistency across views by leveraging epipolar constraints; and finally, a global domain refinement module that performs deep optimization of texture and structure via multi-head self-attention. Experiments on synthetic and real-world LFs demonstrate that the proposed method maintains the consistency of LF structure and simultaneously outperforms state-of-the-art methods in denoising performance.

Snapshot spectral compressive imaging (SCI) is a cutting-edge technique that employs a two-dimensional (2D) detector to capture high-dimensional hyperspectral cubes in a compressed format. The core challenge of SCI systems is how to accurately reconstruct the three-dimensional hyperspectral cubes from the 2D measurements acquired by the sensors. Unlike conventional RGB images, spectral data exhibit strong inter-channel correlations and subtle differences that cannot be ignored. However, existing algorithms are limited due to the scarcity of available compressed spectral imaging data and the overfitting problem in supervised learning, resulting in the loss of fine textures and inter-channel distinctions. Inspired by untrained neural networks and the deep image prior (DIP), we propose a novel approach that integrates deep balanced prior (DBP) and global spectral prior into a framework. This integration results in a self-supervised framework tailored for spectral snapshot compressive imaging reconstruction. Our method surpasses state-of-the-art self-supervised algorithms by approximately 1.8dB when evaluated with synthetic datasets. Extensive experimentation across various datasets demonstrates that our algorithm consistently outperforms existing learning-based methods in reconstructing fine textures from measurements.

Infrared image super-resolution is essential in optical imaging systems to enhance detail recovery for applications such as thermal analysis and night vision. However, it remains challenging due to inherent limitations like low resolution, high noise, and weak contrast. While diffusion-based methods have shown promise, most rely only on low-resolution conditioning, which offers limited high-frequency constraints and lacks adaptive refinement during sampling. To address this, we propose DR-Diff, a spatiotemporal-aware dual residual-guided diffusion model. Our framework integrates a static global residual prior and a dynamic local residual feedback, forming a closed-loop guidance mechanism that supports persistent structure preservation and stepwise local adaptation. Furthermore, a spatiotemporal fusion module dynamically balances conditioning signals via temporal encoding and spatial self-attention, enabling coarse-to-fine reconstruction. Experiments on benchmarks confirm that our approach outperforms existing state-of-the-art methods.