- 學位論文
基於光譜相關性的高品質面掃描式和線掃描式去馬賽克之演算法
High Quality Area Scan and Line Scan Demosaicking Algorithms Based on Spectral Correlation
摘要
隨著高畫質時代(full-HD Generation)的來臨，人們對影像品質的需求越來越高。目前越來越多消費型的成像電子裝置使用單一感光元件，覆蓋一層貝爾排列模式之色彩濾鏡陣列去取樣並記錄場景中相對位置之三原色的像素值，而計算取像過程中丟失的像素值的過程稱為色彩內插演算法(CFA-interpolations)或去馬賽克演算法(Demosaicking Algorithms)。 在本篇論文中，為了減少數位取像裝置的成本及體積並因應高品質的需求，我們提出了基於光譜相關性的高品質面掃描式和線掃描式去馬賽克之演算法。對於面掃描式之成像感測器，我們利用更佳的初始內插演算法以減少影像低頻之誤差，並利用高頻迭代策略有效地得到清晰影像；對於線掃描式之成像感測器，為了有效落實演算法之硬體實現，我們提出利用高斯函數(Gaussian Blur function)分解原始取樣像素以得到平滑層(base layer)和細微層(detail layer)。平滑層採用傳統的色彩內插演算法做影像內插；細微層利用色頻高度相關性作係數迭代。因此，此方法之重建影像比起傳統的演算法，可以獲得較高的PSNR、SSIM值和較佳的視覺品質。
並列摘要
With the advent of the full-HD Generation, human demand for high quality image is growing. Today more and more consumer electronic devices sample and record the nature sense with color pixel values by single sensor covered with color filter array arranged by Bayer pattern. The estimation of lost pixel values during the sensor sampling process is known as Color Filter Array Interpolation (CFA-interpolations) Algorithms or Demosaicking Algorithms. In this paper, in order to reduce cost and size of digital imaging devices, and obtain high quality image, high quality area scan and line scan demosaicking algorithms based on spectral correlation are proposed. For area scan imaging sensors, we utilize better initial interpolations to reduce image errors in low-frequency regions and obtain sharper images by projecting the high-frequency coefficients of each color channels iteratively. For line scan imaging sensors, we propose using the Gaussian Function (Gaussian Blur Function) to decompose the mosaic image into a smooth layer (base layer) and the fine layer (detail layer) for the effective hardware implementation. In base layer, traditional color interpolations are adopted as image interpolation algorithms; in detail layer, perform the coefficient replacements according to the property that for nature images there is a high correlation between color channels. Therefore, reconstruction images, obtained from the proposed method, have higher PSNR and SSIM, and better visual quality than traditional algorithms.