智慧型超解析度系統

Abstract

本論文提出一種智慧型超解析度的方法。此種方法是用於改善圖素型(pixel-based)電腦圖像缺乏解析度獨立性的問題，也就是當影像放大時產生解析度或影像品質下降的問題。一般常用的內插方法通常會使圖像邊緣和細緻內容模糊化。因此，我們發展一種利用訓練模型的內插方法來產生仿真的(plausible)高頻資訊來補償被放大的影像。我們的方法是基礎於樣本應用型超解析度系統(example-based super-resolution)的研究，而其中假設影像的高頻資訊是有關聯於中頻資訊並且與區域影像對比無關。 本研究的目的是改善樣本應用型超解析度系統的缺失和增加它的實用性。我們的改善分為前處理和後處理。在前處理部分，我們利用色彩轉換將彩色影像的照度分離出來進行超解析度處理。如此可減少影像資料量和運算複雜度並且使本系統不受影像顏色影響。在後處理部分，我們分別對區域影像進行特徵抽取並採用模糊推論系統來對那些特徵進行推論而決定出一些加權值來調整被處理影像的最終品質。實驗結果顯示本論文所提出的新方法對於增強影像解析度有令人滿意的效果。

An intelligence-based super-resolution is proposed here. This method is used to solve the problem that pixel-based computer graphics lack resolution independence: they cannot be zoomed much without a degradation of quality. The general interpolating methods for enlarging images usually result in a blurring of edges and image details. Therefore, we develop a training-based interpolation algorithm to create the plausible high-frequency details for compensating the zoomed images. The algorithm is based on the example-based super-resolution [2], with assumptions that the high-frequency band of image is related to the middle-frequency band and independent to the local image contrast. The goal of this research is to improve the defects of example-based super-resolution and increase the practicality of it. The improvements can be divided into preprocessing and post-processing. In the preprocessing, the color space transformation is used to extract the luminance information from color image for super-resolution interpolation. The image data and computation complexity can be reduced; and the system is also independent to the color of image after preprocessing. In the post-processing, the patches of image are analyzed to extract useful features which will be inferred by a fuzzy inference system to decide the weightings used for modulating the final quality of processed image. The experimental results show that the proposed method has good performance on sharpening images.