Dual-Domain Illumination Prior for Low-Light Remote Sensing Image Enhancement

Prior-calibrated spatial-frequency Retinex decomposition for unsupervised low-light image enhancement

DPSF-Net: A Dual-Prior Spatial-Frequency Network for Real-World Remote Sensing Image Dehazing

Real-world remote sensing image dehazing (RSID) remains challenging because atmospheric scattering, spatially non-uniform haze and colour distortion jointly degrade structural and spectral information. Most deep learning methods rely on RGB inputs and spatial-domain feature extraction, which limits their ability to separate global background haze from local surface details. Here, we propose DPSF-Net, a dual-prior spatial-frequency network built on MCAF-Net for real-world RSID. The network uses hazy RGB images and dark channel prior (DCP) maps as joint inputs, allowing physical degradation cues to guide end-to-end feature learning. A spatial-frequency residual interaction block introduces a FourierUnit branch into multi-directional spatial interaction to model large-scale haze components. A prior-guided feature attention module adaptively fuses prior and attention features to reduce colour shift and structural distortion. A selective kernel complementary fusion module screens multi-scale skip features through bidirectional residual complementary gating and selective kernel fusion. Extensive experiments demonstrate that DPSF-Net achieves state-of-the-art performance on the real-world RRSHID remote sensing image dehazing benchmark and remains competitive across multiple synthetic datasets. Moreover, the proposed method strikes a favourable balance among restoration quality, parameter count and computational complexity, supporting the effectiveness of dual-prior spatial-frequency modelling.

Frequency-Spatial Dual-Level Selective Alignment for Domain-Adaptive Object Detection in Remote Sensing

Remote sensing object detection suffers from severe performance degradation under cross-domain transfer, where domain gaps arise from differences in spectral response, spatial resolution, and viewing geometry. Most existing unsupervised domain-adaptive object detection (DAOD) methods pursue cross-domain invariance through feature distribution alignment but encounter two limitations specific to remote sensing: cross-domain appearance variation, where differences in imaging conditions produce divergent visual appearances, and foreground-background imbalance, where targets are sparsely distributed across vast backgrounds and alignment is dominated by background statistics. To address these two limitations, frequency-spatial dual-level selective alignment (FS2A), a framework with two complementary modules, is proposed. At the frequency level, scale-class-conditioned frequency modulation (SCFM) decomposes multiscale features via FFT and selectively modulates the low-frequency amplitude conditioned on scale level and categorical composition, restricting adversarial alignment to domain-variant spectral components while preserving the domain-invariant phase spectrum. At the spatial level, kernel relational distillation (KRD) distills pairwise relational structure from a frozen satellite-pretrained vision foundation model (VFM) in polynomial kernel space, where polynomial kernel functions preferentially concentrate alignment on foreground feature pairs over weakly correlated background pairs. Both modules are decoupled from the detection forward pass, introducing no additional inference cost. Experiments on two cross-domain remote sensing benchmarks demonstrate that FS2A achieves 67.6% mAP50 on xView $\rightarrow $ DOTA, surpassing the state-of-the-art by 2.7%, and competitive results on satellite-to-UAV benchmarks. The code will be available at https://github.com/sparklejojo/FSSA-DAOD

Learning Structural Illumination for Unsupervised Low-light Enhancement

A Relative Illumination Structure Estimation (RISE) framework is proposed that decouples relative illumination structure from absolute exposure and infers it from reliable bright regions, enabling interpretable and robust enhancement.

A region-adaptive hybrid framework for single image dehazing using channel priors and retinex analysis

Haze degrades image clarity, contrast, and fine details, causing perceptual ambiguity. Image dehazing aims to restore visibility, but remains challenging in noisy and complex environments. To address these challenges, prior-based dehazing techniques have been widely explored for effective transmission estimation and scene radiance recovery. The Dark Channel Prior (DCP) uses the minimum intensity across all color channels to estimate transmission in dense haze regions efficiently. This approach works well in non-sky regions with rich textures and darker intensity. However, DCP may not work well in sky regions. On the other hand, the Bright Channel Prior (BCP) uses high-intensity information of pixels to achieve efficient dehazing in bright regions, such as light-dominant regions in the atmosphere, which covers the sky. However, BCP may not work well in non-sky regions. Most existing dehazing algorithms use either DCP or BCP to efficiently recover scene radiance. This results in reduced performance in different regions. To overcome this limitation, a region-based fusion approach is developed to achieve efficient scene radiance recovery from DCP and BCP. The fused image is enhanced using a Retinex-based image decomposition technique, which consists of gamma correction of illumination images and wavelet-based denoising of reflectance images. The performance of dehazing depends on the tuning parameters of the above-mentioned techniques. Also, the ground truth may not be available in real-time scenarios. Therefore, a no-reference (blind) adaptive parameter selection scheme is developed to handle the absence of ground truth in real-world images. The most suitable parameters for DCP and BCP are determined using a grid search guided by the NIQE metric, ensuring improved perceptual quality without requiring reference images. The method is evaluated on I-Haze, O-Haze, Dense-Haze, NH-Haze, and Reside (Reside-6K, RTTS, and UAHI) datasets using SSIM, PSNR, NIQE, and BRISQUE metrics. Experimental results demonstrate consistent improvement across diverse haze conditions. Moreover, the proposed framework operates without any training requirements, removing the necessity for annotated datasets and model optimization processes, all while ensuring low computational complexity, with an average processing time per image is just 0.423 s. Furthermore, the proposed approach demonstrates comparable performance to state-of-the-art methods and is validated through real-world object detection tasks. The source code of the proposed method is publicly available at: RegionBasedAdaptiveHybridImageDehazing.

AASFNet: Amplitude-Aware Spatial-Frequency Network for Low-Light Image Enhancement

Low-light images often suffer from uneven illumination, resulting in reduced brightness, low contrast, and increased noise interference. However, existing enhancement methods frequently lead to over- or under-enhancement, frequency domain distortion, or insufficient noise suppression, which adversely affect both visual quality and the performance of downstream vision tasks. To address these issues, we propose AASFNet, a novel low-light image enhancement network that integrates spatial and frequency domain features through amplitude enhancement and dual-domain fusion. The framework consists of a Frequency-domain Enhancement Network (FreqEnhanceNet) with an illumination-guided Amplitude Module (AmpModule) for adaptive brightness adjustment and a Spatial-Frequency Fusion Network (SpatFreqFusionNet) incorporating a Dual-Domain Fusion Module (DDFModule) for noise suppression and illumination correction via multi-scale feature interaction. This design enables the network to simultaneously adjust global illumination, restore local details, and suppress noise. Additionally, a joint spatial-frequency loss function is introduced, including an illumination-guided amplitude loss and a cosine phase loss, to enhance structural consistency and amplitude fidelity. Comprehensive experiments on the LOL-v2 Real and Synthetic datasets demonstrate that AASFNet achieves competitive or leading performance across multiple key metrics under the evaluated experimental settings, yielding the best or second-best PSNR, SSIM, and NIQE values among the compared methods. Moreover, when applied as a preprocessing module, it achieves improved detection accuracy across multiple low-light benchmarks, with mAP scores of 78.63% on ExDark, 78.2% on DarkFace, and 80.32% on LoLI-Street. These results suggest the potential applicability of the model in real-world scenarios such as autonomous driving and surveillance, although further validation under actual deployment conditions is still required.