多光源水下場景中的影像分割與深度估計：環境調適與強健視覺方法

Abstract

水下視覺感知會因為光線衰減、色彩扭曲與能見度低等問題而面臨眾多挑戰，使得傳統影像處理方法在實際應用中難以維持穩定且準確的辨識能力。為了提升水下場景感知的穩定性，本論文探討了水下影像增強、實例分割與深度預測。首先，我們比較無需訓練與學習型的水下影像增強方法，以改善水下影像品質並對抗視覺退化。接著，我們提出了名為 BARD-ERA 的水下實例分割架構，透過細化邊界感知的結果與設計環境調整機制，提升了在惡劣環境下的目標邊界辨識能力。隨後，我們設計了 SADDER 模組，用以結合分割先驗資訊並改善單張影像的深度預測結果。最後，我們整合實例分割與深度預測，提出一個物體層級的深度感知分割框架，透過物體區域的平均深度計算方式，實現語義與三維資訊的融合。實驗結果顯示，所提出的方法在多個水下資料集上展現優異的準確性與穩健性，具有良好的可解釋性與模組化特性，適用於海洋機器人、生態監測與海底探勘等實務應用場景。

Underwater visual perception presents unique challenges due to issues such as light attenuation, color distortion, and low visibility. These factors significantly hinder the reliability of conventional image-processing techniques when applied to real-world underwater tasks. In response, this thesis explores a series of key modules to enhance the robustness and interpretability of underwater scene understanding. We begin by examining underwater image enhancement, comparing learning-free and learning-based methods to address visual degradation. This is followed by the development of BARD-ERA, an instance segmentation framework that leverages boundary-aware refinement and adapter-based domain tuning to improve object delineation under challenging aquatic conditions. Next, we introduce SADDER, a lightweight depth refinement module designed to enhance monocular depth estimation using segmentation-aware residual learning. Finally, we integrate instance segmentation and depth estimation into a unified depth-informed instance segmentation framework, enabling object-level spatial reasoning by assigning average depth values to segmented regions. Experimental results across multiple underwater datasets validate the effectiveness of the proposed methods, which together contribute to more accurate, interpretable, and modular underwater perception. This research offers practical value for applications such as marine robotics, ecological monitoring, and subsea exploration.