基於最大濾波空間的低光源影像增強方法

Abstract

本篇論文提出一個基於深度學習的低光源影像增強方法，相較於前人的方法直接估計輸出圖片，由於圖片照度相對更加平滑，我們的方法選擇估計輸入圖片的照度，再將原圖除以網路估計的照度來產生最終的輸出圖片；另外，我們基於 Retinex 理論設計了一個針對低光源影像的照度估計方法，並用其當作網路的損失函數之一來幫助其收斂；最後，由於環境與設備上的限制，導致成對資料在取得上較為困難，且在場景及數量上的豐富度較為不足，我們的方法利用非成對資料加上生成對抗網路來處理上述問題。透過與前人的方法比較，發現我們的方法在亮度上有較好的表現，也展示了此方法的可行性。

This thesis proposes a deep learning-based low light image enhancement method. Previous works usually model it as an image-to-image translation problem. Compared with them, our method predicts the illumination of the input image, which is smoother than the image itself. Then we divide the input image by predicted illumination to get the final estimated image. Besides, we devise an illumination estimation method applied to low light images based on Retinex Theory and use the result as a loss function to help our network converge. Finally, it is difficult to get paired data for training because of the limitation on the environment and device. Therefore, researchers generally encounter less scene diversity and insufficient data amount problems when collecting paired data. To tackle the problems mentioned above, we adopt unpaired data and leverage generative adversarial networks. Through the comparison to previous work, we find our method has better performance on brightness, and it shows the potential of our method.