基於色偏補償與自適應背景光估計之水下影像修復研究

Abstract

水下影像修復在海洋探勘、生態監測、水下考古與無人載具等應用領域中扮演關鍵角色，其影像品質將直接影響後續分析與決策的準確性。惟由於水下環境的複雜性，造成影像常出現色彩偏差及清晰度不足等問題。目前，多數水下影像修復技術往往聚焦於單一處理面向，如色彩校正或去霧處理，然而在面對不同水體條件與光照變化時，常出現修復效果不穩定、泛化能力不足等問題，顯示現有方法仍有進一步改良的空間。本研究提出一套創新的水下影像修復技術，整合色偏校正與影像去霧方法。首先，透過色偏因子(Color Cast Factor)分析影像之色偏類型並進行補償處理。接著，採用四元樹搜尋(Quad Tree Search)策略，搭配加權背景光評估函數以精確估計背景光，並以此計算初始透射圖。透射圖進一步結合梯度約束與引導濾波器(Guided Filter)進行改良，以提升邊緣保留能力與抑制雜訊干擾。最終，應用水下成像模型進行影像重建，並透過對比度伸展(Contrast Stretching)強化視覺對比效果。本論文收集超過兩萬張涵蓋不同場景及色偏類型的水下影像，並與七種現有方法進行修復效果比較，以評估所提出的方法。實驗結果顯示，於多種水下場景中，本研究於色偏校正與去霧處理方面皆展現出較高的視覺穩定性。量化評估方面，UCIQE與UIQM兩項指標於多數資料集中皆達到最佳或次佳表現。綜合定量與定性結果，顯示本研究所提出之方法具備良好的應用潛力。

Underwater image restoration plays a crucial role in various applications such as marine exploration, ecological monitoring, underwater archaeology, and autonomous underwater vehicles. The quality of restored images directly affects the accuracy of subsequent analysis and decision-making processes.However, due to the complexity of the underwater environment, images often suffer from color distortion and reduced clarity. Most existing restoration techniques tend to focus on a single aspect, such as color correction or dehazing. Nevertheless, when dealing with varying water conditions and lighting environments, these methods frequently exhibit unstable performance and limited generalization capability, indicating that there is still considerable room for improvement in current approaches. This thesis proposes an innovative underwater image restoration algorithm that integrates color cast correction and image dehazing techniques. First, a Color Cast Factor is used to analyze and compensate for different types of color casts. Then, a quad-tree search strategy combined with a weighted background light evaluation function is employed to accurately estimate the background light, which is subsequently used to compute an initial transmission map. This map is further refined through gradient constraint and guided filtering techniques to enhance edge preservation and suppress noise. Finally, the restored image is reconstructed using an underwater imaging model, followed by contrast stretching to improve visual contrast. This thesis collected over 20,000 underwater images covering diverse scenes and color cast types, and compared the proposed algorithm with seven existing methods to evaluate its performance. Experimental results show that the proposed method demonstrated more stable visual performance in both color correction and dehazing across various underwater environments. In terms of quantitative evaluation, it consistently achieves the best or second-best results in UCIQE and UIQM scores across most datasets. These quantitative and qualitative findings suggest that the proposed method holds strong potential for practical applications.