低複雜度多尺度梯度色彩插補演算法

Abstract

數位相機色彩插補是現在單一感測器數位相機之數位影像處理流程中最重要的部 份之一。近年來有許多基於分類器的色彩插補演算法被提出,傳統分類器色彩插補演 算法雖然再大多數的場合表現的極為優秀,但是存在著因分類錯誤造成嚴重錯色的問 題,因此我們試圖尋找不一樣的解決方案。近年來多尺度梯度演算法 (MSG)是最成功 的演算法之一,其擁有相當卓越的影像品質。多尺度梯度演算法使用權重混合的方式 來取代傳統分類器的作法,雖然這樣的作法可以大幅提高影像品質,但是相對的也大 大的提高計算複雜度。在本篇論文中我們仔細對多尺度梯度演算法進行了分析,研究 其權重設計,試圖尋找一種適合的替代設計,能在維持多尺度梯度演算法的影像品質 的情況下,降低其計算複雜度。本方法主要特點在於,合併了兩階段綠色平面插補的 權重設計。在初始插補綠色平面階段以及優化綠色平面階段能使用同一組權重矩陣設 計,有效的降低計算複雜度。實驗結果證明我們不僅在 PSNR 能取得與多尺度演算法 接近的表現,在較為接近人類視覺的 S-CIELab 也能有接近多尺度梯度演算法的表現。 在執行時間上,比起多尺度梯度演算法有明顯的改進,其方法一執行時間約減少 36.37%,而方法二則減少了 41.52%。Demosaicing of the color filter array (CFA) is one of the most important parts of the image processing pipeline for single sensor digital cameras. In the recent years, many classifer-based demosaicing algorithms have been proposed and achieved excellent performance in most situations. However, occasionally these algorithms generate results with color artifacts due to wrong directional leads. Recently, the multiscale gradients (MSG) algorithm was reported, which has the best reported image quality to this date. The MSG algorithm used adaptive weight design to replace the conventional classifer-based design. This design boosts the image quality; however, it made a trade-off with computational complexity. In this thesis, modification and improvement to the MSG algorithm by low- complexity multiscale gradients algorithm is attempted such that the computational complexity is significantly reduced while maintaining the high image quality possessed by the original MSG algorithm. Experimental results show that the proposed algorithm not only maintains peak signal-to-noise ratio (PSNR) and S-CIELab. Two methods were proposed in this thesis and the speeds compared to the original MSG algorithm were increased by 36.37% and 41.52%, respectively, which is a significant improvement.