應用擴散式生成對抗網路於複雜光線條件下的影像曝光修正

Abstract

曝光修正的任務是將一張存在過度曝光或曝光不足的影像修正，使畫面中物件能夠清晰可見。過去的工作偏向處理整體畫面為過度曝光或是曝光不足的現象，例如，低光增強。但在不同光線存在的情況下，一張影像經常會有同時有局部過度曝光及局部曝光不足的情況。過去的工作無法用來解決兩種問題同時存在的情況，但在實際應用中，影像中常常同時存在過度曝光以及曝光不足這兩種情況。為了解決一張照片同時有過度曝光以及曝光不足的問題，本研究引入一個生成擴散式對抗網路的方法，透過向辨別器添加高斯雜訊使辨別器可以訓練更好，使得能夠生產更好的曝光修正模型。同時，本論文引入兩種不同的曝光修正損失，以確保生成對抗模型能夠擁有更加穩定的訓練。此外，本論文還加入基準影像與輸入影像都相同且都是曝光正確的成對影像，增加模型對於曝光正確影像的理解。本論文的實驗結果表明，本論文提出的方法能有效應對同時擁有曝光不足及過度曝光的挑戰，並在三種不同的評估標準上有不錯的表現，並同時在曝光正常的資料集也不會產生不合理的曝光調整，未來的研究將拓展這一方法，處理更複雜多樣化的影像環境。The task of exposure correction involves rectifying an image that suffers from either overexposure or underexposure, ensuring that objects in the scene are clear and visible. Previous approaches predominantly focused on addressing overall issues of overexposure or underexposure, such as low-light enhancement. However, in scenarios with varying lighting conditions, an image often exhibits both localized overexposure and underexposure simultaneously. Past methods were insufficient for handling situations where both problems coexisted. In practical applications, images frequently exhibit both overexposure and underexposure concurrently.To address the challenge of simultaneously dealing with overexposure and underexposure in a single photograph, we introduce a generative adversarial network (GAN) approach. We incorporate Gaussian noise into the discriminator to enhance discriminator’s training, enabling the generation of superior exposure correction models. Simultaneously, we introduce two distinct exposure correction losses to ensure the stability of training in the generative adversarial model. Additionally, we include pairs of reference images and input images that are both correctly exposed, enhancing the model's understanding of correctly exposed images. Our experimental results demonstrate the effectiveness of our proposed method in tackling the challenges of simultaneous underexposure and overexposure. The approach performs well across three different evaluation standards and does not produce unreasonable exposure adjustments in datasets with normal exposure. Future research will expand on this method to address more complex and diverse image environments.