基於改進視覺信號保真度之影像品質評估

Image Quality Assessment Based on Improved Visual Information Fidelity

謝昌利(Chang-Li Hsieh)

指導教授 ： 郭天穎

國立臺北科技大學/電資學院/電機工程系研究所/碩士(2014年)

摘要

視覺信號保真度(Visual information fidelity，VIF)為一影像品質評估法，其考慮了人類視覺系統(Human Visual System，HVS)特性以做到主觀評估參考與失真兩張影像間之差異程度。
本文提出基於視覺信號保真度的評估方法做改進，讓影像評估同時著重於包含物件顯著區域和失真較嚴重的視覺敏感區域。我們還透由修改視覺信號保真度演算法，改善視覺信號保真度高估與低估之情形。
當失真嚴重且集中時，人眼對影像品質會驟降之情形，視覺信號保真度容易高估品質評估，我們利用衰減信號特徵圖上考慮信號衰減較強之集中性來判斷，以調高失真影像在視覺信號保真度的人類視覺通道模型中之視覺雜訊變異量，以降低影像的評估分數，符合人眼特性。
另外在人眼比對兩張極相似的影像時，視覺信號保真度會低估評估分數的問題，所以我們藉由比較兩張影像在高斯尺度混合(Gaussian Scale Mixture，GSM)模型上之誤差，讓誤差低於門檻時提高評估分數。
最後由於視覺信號保真度利用灰階影像評估互信息量，沒有考慮對彩色影像資訊，所以讓演算法加入色差演算法CIEDE2000，讓評估著重於顏色差異較大的局部區域且提升視覺雜訊的變異量，來彌補視覺信號保真度的不足。

關鍵字

影像品質評估 ； 人類視覺系統 ； 主觀評估 ； 視覺信號保真度 ； 高斯尺度混合 ； 色差演算法CIEDE2000

並列摘要

Visual information fidelity (VIF) is a method of image quality assessment (IQA), which exploits the characteristics of the human visual system (HVS) for subjectively evaluating the difference between the reference and distorted images.
This work proposed an improved VIF-based IQA by taking the visually sensitive region with salient objects and with severe distortion into account to enhance VIF performance. In addition, we also consider both the over and under-estimated cases of VIF.
When the distortion is harsh and centralized on a spot, VIF will overly assess as human vision is very sensitive to that degradation. We check if the distortion operator between two images are strong and spatially centralized, we can increase noise variance of the HVS channel model in VIF for distortion images, to lower VIF score in accord with the HVS.
On the contrary, VIF has the under-estimation problem when both images are very similar. So we check the differences in both images using the Gaussian scale mixtures (GSM) model. If difference is smaller than the threshold, the score of the image quality is promoted.
Since VIF calculates the mutual information by the gray-level images without considering color, we incorporate the color difference model of CIEDE2000 to VIF. Thus, our improved VIF can address the region and the noise variance with the area of large difference in color.

並列關鍵字

Image quality assessment ； human visual system ； subjective assessment ； visual information fidelity ； Gaussian scale mixtures ； CIEDE2000

參考文獻

[19]徐詩媛，考慮感知程度之影像品質評估，碩士論文，國立臺北科技大學電機所，臺北，2013。

[2]Z. Wang and A. C. Bovik, Modern Image Quality Assessment, CA: Morgan & Claypool Publishers, 2006.

[3]Luo MR, Cui G, Rigg B. “The development of the CIE 2000 colour-difference formula: CIEDE2000,” Color Research & Application, vol. 26, no. 5, Oct. 2001, pp. 340-350.

[4]H.R. Sheikh and A.C. Bovik, “Image information and visual quality,” IEEE Trans. Image Process., vol. 15, no. 2, Feb. 2006, pp. 430-444.

[5]M. Antonini, M. Barlaud ,P. Mathieu and I. Daubhechies, “Image coding using wavelet transform,” IEEE Trans. Image Process., vol. 1, no. 2, Apr. 1992, pp. 205-220.