運用視覺焦點分析與最低可視別雜訊的影片品質評估法

Abstract

近年來，由於網路發達與硬體上的演進，人類在數位影像的需求越來越大，對於品質的要求也跟著科技的無遠弗屆相對的提升。不管是在研究或是應用上，視覺品質量測(Quality Assessment)都代表著一個引領的地位：研究者會依據評比而調整內容，或是在開發應用時也會根據此評比數據而調整演算法。因此，使用人眼的主觀來評斷一個影片的優劣，或是單單使用影片差異的量化去對影像做評比，前者對於現今我們所需求的，就顯的相對的重要。 傳統上我們通常使用訊雜比(PSNR)作為影像品質評定的標準，但是經實驗證實，此方法與人眼視覺系統有著相當低的關連性。因此，許多的相關的研究如雨後春筍般相繼冒出。此篇論文即是介紹一個創新、直觀、並且容易執行的影像品質評定方法，以模擬人類視覺上對於各種受損影像的反應程度。 我們利用”最低可察覺損害”(Just-Noticeable Distortion)的技術，加上人眼視覺上對於不同部位會有不同權重的特性，去模擬視覺上對於受損影像的評估。除此之外，為了能夠更接近人類感知系統，我們再加上兩個特徵項，用以加強演算法的穩定度: 1. 時間軸變化量(Temporal information fidelity). 2. 塊狀雜訊偵測(blocking artifact detection)。根據實驗結果，我們可以發現，此三項特徵項的組合，相對於其他演算法，較符合人眼視覺的評斷機制。

Traditionally, peak-signal-to-noise ratio (PSNR) has been used to represent the quality of numerous videos and images. However, it has been found that PSNR has a poor correlation with the true feelings of the human eyes. Therefore, due to the rapid increase in the usages of digital videos, we need a more explicit quality assessment method to access the scores of videos, which can not only measure the signal impairment amount in different distortion types of videos but also take the human visual system into consideration. This thesis combines two models to find the distortion visibility of human eyes of videos. One is the just noticeable distortion (JND) approach, and the other is the visual saliency model. The former can estimate the maximum distortion that the human visual system cannot perceive, and the latter can predict human fixations on video frames. Besides, to modulate the perceptual quality to be more consistent, our proposed model incorporates the block fidelity and the temporal information fidelity which have been considered as the most noticeable components of video distortions by human perception. The proposed method has been tested on 90 videos with a variety of distortions, such as MPEG-2 compression, Intra-coded frame loss, packet loss and H.264. compression, and has been shown to give a significantly better correlation to the subjective quality score than PSNR and other quality measurement.