使用深度學習模型進行天空去噪

Abstract

「影像去噪」旨在消除由電子傳輸、圖像信號處理和圖片壓縮產生的噪聲，是電腦視覺領域的一個重要研究主題。它可以幫助提高影像品質，且能增進後續其他任務的表現，像是：影像分類、物件偵測、物件分割等。「基於影像語意的影像增強」指的是將照片根據語意劃分為多個區域（例如：天空、樹、人），並根據每種區域不同的特性施以專門的降噪器或美化器。

在這篇論文中，我們專注於使用深度學習網路做天空區域的去噪，並以此為目標進行以下三步驟。首先，調查使用深度學習進行影像去噪最核心的兩個議題：網路架構設計與生成噪聲影像。我們調查了近幾年的研究是如何解決這兩個問題，並對其做分類與歸納。再者，為了選出最有潛力的降噪器，我們把來自十篇不同研究的降噪器測試在天空的噪聲圖片上，並採用表現最好的網路進行後續的重新訓練。最後，為了提高降噪器的表現，我們嘗試了不同的噪聲合成步驟，並選用效果最好的來訓練最終的模型。

Image denoising aims to remove the noise from degraded observations generated from the electronic transmission, image signal process pipeline, and compression. Image denoising is a prevalent research topic in computer vision, and assists in enhancing image quality and improving the performance of high-level computer vision tasks, such as image classification, object detection, and object segmentation.Segment-based enhancement divides a photo into semantic-based regions, and enhances each region with different beautifiers and denoisers according to the semantic domain knowledge.

In this paper, we focus on sky denoising with deep learning approaches. To achieve this goal, we execute the following three steps: First, we survey and summarize how the widely-used and state-of-the-art researches tackle the two critical components of image denoising: model architecture and noise data synthesis. Secondly, we execute a pilot study on ten different denoisers, which have released the pre-trained models, and we retrain the best-performing model on our sky dataset. Finally, we try different procedures of noise synthesis to improve the retaining performance and select the best one to train our final model.