使用原始圖檔於低光環境物件偵測

Abstract

物件偵測被廣泛運用在監視器、自動駕駛等領域。近來來受益於深度學習，物件偵測技術有突破性的進展，然而低光環境會對此技術造成的不良影響，使得現存在JPG圖片上的方法容易失敗，導致既有的應用的成效不彰。為此我們提出第一個在RAW圖片上的端對端訓練物件偵測模型。並且基於影像訊息處理流水線提出新穎的架構—— ConvISP，透過預測影像訊息處理的參數，來達到最適化低光環境的影像處理流程。廣泛的實驗結果說明此架構使得物件偵測在不同程度低光環境仍能成功運作。

Object detection has been applied in many areas, including security, surveillance, automated vehicle systems, and so on. In recent years, the deep learning-based approaches for solving objection detection have achieved great success. However, the performance deterioration under low-light conditions, which makes the existing algorithms trained on JPG images prone to fail, is inevitable in real-world applications (e.g., adverse weather conditions). To this end, we propose the first end-to-end trainable object detection model on RAW images. Inspired by the Image Signal Processing (ISP) pipeline, we design one novel component called \emph{ConvISP}, which aims at predicting ISP parameters. Extensive experimental results demonstrate that the proposed framework works well under low light conditions in different degrees.