基於結構相似度之改良式彩色影像品質評估法

Abstract

如何探討影像的品質是影像處理中很重要的一件事。最近許多研究如何客觀計算出影像品質的演算法一一的被提出。傳統的MSE或PSNR是以兩張圖片絕對的差異為基礎去計算品質，計算出來的結果常常和人眼主觀的評價不同。因此，以人類視覺系統(HVS)為基礎的影像品質評估法漸漸取代了它。 我們介紹了基於結構相似性為基礎的影像品質評估法－SSIM 。根據原本的圖片還有失真的圖片之間結構所提供的資訊對於人眼的相似度，我們計算出了失真圖片的品質。SSIM相對於PSNR或MSE明顯的比較接近人類主觀評價的結果。許多以SSIM為基礎的改進法也一一的被提出，這些改進的方法考慮了更多人眼的特徵或者把單解析度的SSIM發展到多解析度的SSIM，這些改進法提供了更精準且更彈性的結果。 我們生活在一個色彩繽紛的世界中，大部分看到的圖片或影片也都是彩色的。然而，大部分的影響品質評估法卻都是實施在黑白圖片上。彩色也是一個很重要的特徵，所以我們提出了一個影像品質評估法來處理彩色的圖片。加入了顏色的特徵後，我們可以看到我們提出的演算法表現的比原本的更接近人眼的主觀感覺。我相信發展出了一個接近人眼的影像品質評估法，對於應用在影像處理上，會非常有幫助。

Image quality assessment (IQA) plays an important role in various image processing applications. In recent years, the research of objective image quality metrics has been developed widely. Human Visual System (HVS) based image quality assessments take the place of simple methods based on absolute difference between pixels of two images (such as MSE and PSNR). In other words, we hope that result of objective evaluation is consistent with human subjective opinion. We introduce Structural SIMilarity (SSIM) based full-reference image quality measurement. The score of SSIM depends on perceptual similarity of structural information between reference image and distorted image. SSIM has better performance than PSNR (or MSE). Many improved methods based on SSIM are proposed one by one. For example, those methods consider the human emphatic features. Multi-scale SSIM provides more flexibility than single-scale SSIM. Those improved algorithms provide better performance than original SSIM. Most image quality assessments are applied in gray-level images. However, color image is widely used in recent years. Color is also an important feature of an image. It is necessary to develop image quality assessment to perform in color image. We propose an improved color image quality algorithm to deal with color images. In our simulation, we can see the better result of algorithm after considering color feature of image. The result after adding our proposed algorithm is more consistent with human subjective perception than original algorithm.