利用圖像增強及邊緣精鍊達到影像放大之研究

Abstract

利用單一影像做到超解析度影像重建( image super-resolution reconstruction )是近幾年來在電腦視覺、圖學領域上熱門的研究領域。在科技發展快速的時代，許多行動裝置提供的低解析度影像已不符合大眾需求，因此如何利用一些現有的低解析度影像分析其資訊進而做到無失真、可信並快速的超解析度影像放大成為一大課題。 這篇論文中我們提出一個新的方法，結合影像增強、相似區塊填補和利用邊緣資訊內插法去重建出一張粗略放大影像。利用影像增強，我們可以保留更多邊緣資訊避免過度平滑並且讓它看起來更接近原始圖。接著利用邊緣區塊相似性去填補讓邊緣連接處較平順，最後再用邊緣資訊像是梯度、向量…等內插，使整張圖概略放大。最後在利用顏色、距離、梯度、梯度值去做精鍊，在合適的時間內保留原始圖片特徵做到少失真的超解析度影像重建。

Single image super-resolution has recently become a hot topic in computer vision and computer graphics communities. This technology had been applied on various media devices in our daily life. The problem how to enlarge images without artifacts and in real time is the core of super-resolution we need to solve. We propose a new method combining image enhancement and interpolation according to edge gradient information. We enhance the original input image at first. This can avoid oversmooth and let the image looked more like the ground truth image. Then we use edge direction and color information to interpolate the unknown pixels, this way can help retain the edge structures roughly. After rough image interpolation we refine the images making it accurate. This method preserves the original features and the nature texture and the whole system can be executed in adaptive time.