基於生成式對抗網路與增強模塊之水下影像強化研究

Abstract

光在水中傳播十會受到水中濃密混濁介質與水中普遍存在的懸浮微粒影響，容易造成光散射、吸收等現象，導致水下能見度降低，並使水下影像產生霧化以及對比度不足的問題。另一方面，不同顏色的光因為波長不同，而逐漸被水吸收，造成水下影像呈現偏藍或偏綠的色偏現象。近年來，除了使用傳統方法，越來越多學者嘗試使用機器學習的方法來修復水下影像。本研究主要探討使用生成式對抗網路修復水下影像之應用。我們利用現有水下影像資料以及本研究所設計的合成水下影像作為訓練集，透過U型網路、卷積注意力模塊以及金字塔池化模塊來開發一個水下影像修復的模型。本論文收集多種不同場景和色偏的水下影像，並使用這些影像來評估本研究所提出的方法。實驗結果顯示，在各種水下場景的修復中，我們提出的方法相對於現有的許多方法具有更穩定的色偏修正和去散射能力，表現出在多種水下影像修復應用中的潛力。

Underwater environments alter the appearance of objects, which makes underwater mage restoration a challenging problem due to multiple distortionsDegradation in image information is primarily caused by the effects of light scattering, wavelength-dependent color attenuation, and object blurrinessMachine learning has become a popular alternative to traditional methods in recent yearsThis study focuses on applying an adversarial network to restore underwater imagesOur model is developed based on a combination of existing and synthetic underwater images as a training set and constructed with U-net, convolution block attention modules, and pyramidal poolingWe evaluated the proposed method using a wide variety of underwater images with various scenes and color shiftsExperimental results suggested that our network outperformed many existing methods in terms of stable color shift correction and de-scattering capability, which demonstrated its potential in many underwater image restoration applications.