採用投映視窗之乏晰類神經網路在影像內插與重呈現上之應用研究

Abstract

影像內插的目地是將低解析度的影像還原為高解析度的影像，此內插的功 能在許多的應用上扮演相當重要的角色。影像內插的過程，可被視為應用 投射函式將輸入的影像轉換到輸出影像，此函式稱為內插函式。過去的研 究所提出的方法都是給予一指定的函式模式。本論文提出一種新的方法利 用類神精網路並具有學習的特性，可學習出一可適性的高階內插函式。這 種具有學習能力的特性有別於傳統的方法。可是，在實際的應用上，這個 問題開始時只有輸入的部份影像是已知的，而放大後的影像尚無法得知。 因此很難取得最佳的學習樣本。本論文於是採用由投影取得影像的模式， 把它利用在學習樣本的取得；而取得樣本的方法主要則是利用一視窗掃描 輸入的影像，由此而影像內插的工作就可視為投影取得的影像的反轉換操 作。實驗結果證明此種新的方法可以得到適當而不錯的結果。 Image interpolation for reconstructing images from low resolut- ion to high resolution is an important processing step for many applications. The image interpolation process can be viewed as a transformation function, called interpolation function, from input subsampled image to interpolated image. During the past years, a lot of approaches using some pre-specified and non- adaptive function models are proposed. In this thesis, the method based on neural network with learning property is different from the conventional approaches. Because that the problem input is the subsampled image only and the target output is unknown in the real-world application, it is difficult to decide the optimal sample set for neural network training. However, the projection model of image acquisition is proposed and applied to the generation of training samples with a window scanning in the input image. Thus, the image interpolation process can be viewed and models as an inversion of image acquisition. Based on this idea, our experimental results demonstrate that our proposed methods are proven to be useful and successful in solving this problem .