DCP-Guided One-Shot Remote Sensing Image Dehazing via Self-Supervised Redegradation Model
Jianchong Wei, Yan Cao, Dongying Chen, Pingping Chen, Shahid Mumtaz, Hadeel Alsolai, Chengbin Chen
Fujian Jiangxia University Fuzhou University Nottingham Trent University Princess Nourah bint Abdulrahman University
阅读操作
确认中在文库中上传 PDF 后可生成中文音频讲解。
摘要与影响
Haze often degrades remote sensing (RS) imagery, reducing contrast and obscuring structural details, which undermines the accuracy of downstream vision tasks. While recent one-shot dehazing methods offer improved generalization by training on a single image, they suffer from prohibitively slow convergence due to unstable optimization from random initialization. To address this, we propose a dark channel prior (DCP)-guided one-shot dehazing framework that integrates a non-trainable DCP module with a redegradation haze imaging model for self-supervised optimization. The key idea is to use the DCP module to provide stable initial estimates of the atmospheric light and transmission map, while two UNet-based subnetworks refine the dehazed image and transmission map under redegradation-based physical consistency constraints. An adaptive warm-up strategy is further introduced to prioritize DCP-regularized learning in the early stage and then gradually transition to full joint optimization as the network predictions stabilize. Extensive experiments on synthetic and real-world RS datasets show that, although our method does not achieve the highest Peak Signal-to-Noise Ratio (PSNR) compared with fully supervised Convolutional Neural Network (CNN)-based models trained on large paired datasets, it attains competitive Structural Similarity Index Measure (SSIM) and clearly better perceptual quality among one-shot and prior-based baselines, while significantly reducing training iterations compared with existing one-shot methods.Moreover, evaluations on a road extraction task demonstrate that the dehazed outputs notably enhance segmentation accuracy and robustness, confirming the practical value of the proposed framework for high-level RS vision applications where paired training data and long per-image optimization are impractical.
逐年被引趋势
关键指标
同类平均 = 1
同领域 · 同年份 · 同类型
Google Scholar 与 OpenAlex 的被引统计范围不同,数值存在差异属正常。
AI 辅助阅读
依据:摘要
可就本文提问;依据不足时会说明。
学术脉络
学科主题
计算机 / AIImage Enhancement Techniques
Image and Signal Denoising Methods · Advanced Image Processing Techniques