基於細緻穿透圖與局部大氣亮度估測的單一影像去霧神經網路

Abstract

拍攝數位影像時，相機的感光元件所接受到的光源，除了來自所拍攝的物體與景象所反射光線外，也會接受到來自空氣中懸浮微粒所反射的光線。霧，便是一定濃度的懸浮微粒聚集在空氣中，吸收與反射光線所形成的自然景象。在起霧情況下所拍攝的照片，有著能見度低、對比度低、色彩飽和度低等特性。這樣的特性，廣泛影響各種電腦視覺相關任務，例如:邊緣檢測、提取圖片特徵、物件分類、監控系統等等。因此，若能將有霧的圖片還原得到一張清晰的影像，會有助於改善電腦視覺任務的判別與效能。 基於物理模型，本篇論文提出了一個深度學習的演算法，使用DenseNet提取霧的多個特徵，並利用兩個解碼器共同估計穿透圖(transmission map)和大氣光(atmospheric light)，最後通過細緻化模組(refinement module)獲得去霧的清晰影像。在合成訓練資料的步驟中，考慮到真實場景下有霧圖片對比度較低的特性，因此加入降低對比度的步驟。同時，為了使模型能更精準的預測場景的穿透度，對產生穿透圖的場景深度進行細緻化處理。最後，為了提升模型的泛化能力考慮了涵蓋不同濃度的室內與室外場景。在訓練過程中，為了增強對遠方場景的除霧能力，損失函數使用衍生自穿透圖的權重數值進行均方誤差計算。在實際運用上，觀察到局部區域所估測的穿透圖較為精細，並考量到真實場景中的大氣光可能存在局部區域的不一致性，因此設計了一個局部估計方法，用於估測出較精細的穿透圖與局部區域大氣亮度，以提高在現實生活中的適用性。

When taking a digital image, the light received by the camera photo sensors includes not only the reflected light from the shooting scene but also the scattering light reflected by the particles suspended in the atmosphere. Haze is the weather phenomenon resulted from the extremely small particles in the air absorbing and scattering the light beam. Digital images captured under hazy circumstances will possess with the characteristics of poor visibility, lower contrast, and reduced color saturation. These characteristics widely affect various computer vision tasks, such as edge detection, feature extraction, object classification, and monitoring systems, etc. Therefore, recovering a clear and pleasing image would significantly improve the performance of the computer vision algorithms. Based on the atmospheric light model, the proposed algorithm utilizes DenseNet to extract multiple features of hazy image, employs two decoders to jointly estimate the transmission map and global atmospheric light, and eventually obtains the final result through a refinement module. The training data is synthesized with a meticulous procedure that considers the contrast of hazy image, the exquisiteness of transmission map, and the density of variant hazy scenes. In order to strength the dehaze ability on remote scene, the loss function applies weighted mean squared error computation with weight derived from the transmission map. Observing the estimated local transmission map has finer prediction, and considering that the atmospheric light in real scene may exist inconsistent in local regions, a local estimation method is designed to estimate finer transmission map and local atmospheric light to enhance the applicability of dehazing in real life.