基於Retinex的水下影像增強方法;Retinex-Based Underwater Image Enhancement

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/98378

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98378

Title:	基於Retinex的水下影像增強方法;Retinex-Based Underwater Image Enhancement
Authors:	鍾意涓;Chung, Yi-Jyuan
Contributors:	資訊工程學系
Keywords:	水下影像增強;Retinex理論;Cross Attention;色彩衰退先驗;Underwater Image Enhancement;Retinex Theory;Cross Attention;Color Attenuation Prior
Date:	2025-07-28
Issue Date:	2025-10-17 12:42:21 (UTC+8)
Publisher:	國立中央大學
Abstract:	水下影像增強(Underwater Image Enhancement, UIE)因其在探索未知水域、海洋資源調查與水下自動化應用上的重要性，成為近年熱門研究課題。然而，水下影像經常面臨低可視度、色彩失真、動態模糊與光照不均勻等問題，傳統方法因過於依賴特定環境參數而限制了泛化能力與效果。本研究結合Retinex理論與深度學習技術，提出一種基於雙分支網路架構的水下影像增強方法，其中亮度與色彩資訊透過獨立分支先分別處理，後續透過新設計的Cross Attention模組進行色彩、亮度資訊交互與融合。亮度分支利用光照強度特徵強化細節表現；色彩分支則採用色彩衰退先驗(Color Attenuation Prior, CAP)進行預處理，有效矯正水下嚴重的色偏現象。本研究所提出的交叉注意力機制藉由多頭注意力方法實現雙向資訊流動，大幅提升了影像的整體視覺品質，根據實驗結果顯示，在LSUI和UIEB水下影像資料集均展現卓越的性能，於PSNR、SSIM和UIQM等評估指標明顯優於現有代表性方法，證實其泛用性與有效性。 ;Underwater Image Enhancement (UIE) has become a prominent research topic in recent years due to its significance in exploring uncharted waters, marine resource surveys, and underwater automation applications. However, underwater images often suffer from poor visibility, color distortion, motion blur, and uneven illumination. Traditional methods tend to rely heavily on specific environmental parameters, which limits their generalization ability and effectiveness. In this study, we propose a novel underwater image enhancement method that integrates the Retinex theory with deep learning techniques, utilizing a dual-branch network architecture. The proposed framework processes luminance and color information separately through independent branches and incorporates a newly designed Cross Attention module to enable effective interaction and fusion of luminance and color features. The luminance branch enhances fine details by leveraging illumination intensity features, while the color branch adopts the Color Attenuation Prior (CAP) for preprocessing, effectively correcting severe color casts in underwater scenes. The proposed cross-attention mechanism facilitates bidirectional information flow through a multi-head attention strategy, significantly improving the overall visual quality of the enhanced images. Experimental results on the LSUI and UIEB underwater image datasets demonstrate that the proposed method achieves superior performance compared to existing representative approaches, as evidenced by notable improvements in PSNR, SSIM, and UIQM metrics, thereby validating its generalization ability and effectiveness.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	8	View/Open

社群 sharing

Loading...