基於雨嵌入一致性和注意力機制之單張影像去雨

Abstract

由於數位媒體的快速發展，影像處理的技術越來越受到人們的重視。不過由於影像資料之來源非常廣泛且品質難以控制，往往會有不同種因素的干擾，包括障礙物、光源、天氣等等，造成影像品質過低，可能會使其相關應用之效能大打折扣，甚至毫無用途。因此，為了解決這些難題，人們投入數位影像品質回復或強化的研究，在近些年來取得明顯的提高影像判讀性及可視性，還能幫助提高物件偵測的準確率。而在我們日常生活中，下雨是最常出現的情況，造成不管是拍攝影像或影片，都會因為雨水而造成影像不清晰。在目前現有的研究方法裡，有使用深度學習、多尺度、Transformer模型等影像去雨方法。其中在使用編碼器解碼器的去雨方式裡，通常是根據輸入的有雨影像來預測雨層。因此，編碼器解碼器的網路架構引起了廣泛的關注。但由於在編碼器階段需要提取影像裡有雨的特徵，而在提取的效果及精確度就很重要。為了解決這個問題，許多論文會加上各種模塊來提升提取的效果。為了解決上述問題，本篇論文提出一個編碼器解碼器網路架構， 並且加上注意力模塊，使其在編碼器階段可以提取更多更準確的有雨特徵，且在編碼器解碼器裡常用的跳躍連接也改成注意力機制的模塊，以讓編碼器提取的特徵可以加強傳遞，使得解碼器可以更為準確預測雨層。在實驗階段，我們使用了多個知名影像資料集，包括Rain100H、Rain100L以及Rain800來訓練及測試所提出的網路架構效果。

Because of the rapid development of digital media, people pay more and more attention to image processing technology. However, due to the wide range of sources of image data and the quality is difficult to control, there are often interferences from various factors, including obstacles, light sources, weather, etc., resulting in low image quality, which may greatly reduce the performance of related applications, or even Useless. Therefore, in order to solve these problems, people have invested in the research of digital image quality restoration or enhancement, which has significantly improved image interpretation and visibility in recent years, and can also help improve the accuracy of object detection. In our daily life, rain is the most common occurrence, and whether it is shooting images or videos, the images will be unclear due to rain. Among the existing research methods, there are image deraining methods using deep learning, multi-scale, and Transformer models. Among them, in the rain removal method using an encoder-decoder, the rain layer is usually predicted based on the input rainy image. Therefore, the network architecture of encoder-decoder has attracted extensive attention. However, since the feature of rain in the image needs to be extracted in the encoder stage, the effect and accuracy of the extraction are very important. In order to solve this problem, many papers will add various modules to improve the extraction effect.In order to solve the above problems, this paper proposes an encoder-decoder network architecture, and adds an attention module, so that it can extract more and more accurate rainy features in the encoder stage, and is commonly used in encoder-decoder The skip connection is also changed to an attention mechanism module, so that the features extracted by the encoder can be strengthened, so that the decoder can predict the rain layer more accurately. In the experimental phase, we used several well-known image datasets, including Rain100H, Rain100L and Rain800, to train and test the effect of the proposed network architecture.