水下影像色彩還原之物距精度影響及最佳參數配置分析

Abstract

透過物理模式進行水下影像色彩還原的方式，可以藉由估計水下影像成像原理中的環境參數，反向推算出場景在空氣中應呈現的色彩。然而，此方法必須建立在環境參數具有足夠精度的前提下。其中，物距作為影響水下影像成像效果的重要因素，其精度對色彩還原成效的影響分析尚付之闕如。因此，本研究分別於模擬實驗與實際實驗中對場景進行控制以獲取高精度物距資料，並在引入不同等級的隨機誤差後進行影像色彩還原，分析物距精度對還原成效的影響，進一步找出達到設定之還原品質閾值所需的物距精度。在實際實驗中，因缺乏控制點及可作尺度參考的人工物體，因此於場景中放置已事先量測點位坐標的塑膠框，作為控制基準以獲取物距資料。 為比較不同水域環境對精度需求的差異，本研究在模擬實驗中建立了從「混濁至清澈」及「淺水至深水」等不同環境條件下的模擬影像，並於實際實驗中將實驗水域分類，找出與其相應在模擬實驗中最相近之水域，再比較兩者之實驗結論，以作為針對多樣水域狀況的抽樣驗證。除此之外，本研究亦針對不同水域環境，分析使色彩還原成效最佳化所需的參數配置，提供後續研究參考。最後，本研究亦探討了相同等級物距誤差對同一影像中不同物距位置的場景所造成的差異影響，以及比較以原始影像與還原影像產製的點雲網格模型，驗證色彩還原程序有助於提取場景資訊並提升水下攝影場景重建的幾何完整度。

Physical-model-based underwater image color restoration can reconstruct the original colors of underwater scenes in air by estimating the environmental parameters involved in underwater imaging formation. However, the quality of the obtained environmental parameters must be sufficiently high. Among them, object distance plays a crucial role in affecting underwater imaging but has not yet been analyzed in terms of its accuracy impact on restoration performance. Therefore, this study establishes experimental settings in both simulation and real-world environments. By controlling the scene, high-accuracy object distance maps are acquired, and random errors of varying levels are introduced into the object distance maps before applying the color restoration process, in order to analyze the effect of object distance accuracy on restoration performance and determine the required accuracy for achieving restoration threshold. In the real-world experiments, due to the lack of control points and artificial objects that could serve as scale references, plastic frames with pre-measured coordinates were placed in the scene to enable control and obtain object distance maps. To investigate the differences in accuracy requirements across different underwater environments, simulated underwater images under various conditions, ranging from “turbid to clear water” and from “shallow to deep water,” were generated for analysis. In the real-world experiments, the aquatic environment was classified and compared with its closest underwater environments from the simulation experiments, serving as a sampling validation of the diverse water types analyzed in the simulations. Furthermore, this study also analyzes the optimal parameter configurations required to achieve the best restoration results across different environments, providing references for future research. In addition, the impact of identical levels of object distance errors on scene restoration performance at different distances within the same image is examined. Finally, by comparing mesh models reconstructed from original and restored images, the study demonstrates that color restoration facilitates the extraction of scene information and improves the geometric completeness of underwater scene.