摘 要
圖像質(zhì)量評(píng)價(jià)對(duì)圖像處理起著重要的指導(dǎo)作用。在很難得到原始圖像的情況下，圖像的無(wú)參考評(píng)價(jià)方法受到了普遍的重視。圖像為人服務(wù)的本質(zhì)決定了它的評(píng)價(jià)結(jié)果必須與人視覺(jué)系統(tǒng)特征相吻合，因此，圖像質(zhì)量評(píng)價(jià)就是設(shè)計(jì)有效的算法以得出與人主觀感知相一致的評(píng)價(jià)值。JPEG圖像依然是當(dāng)前網(wǎng)絡(luò)和數(shù)據(jù)庫(kù)中應(yīng)用最為廣泛和最受歡迎的圖像格式之一。據(jù)此，本論文對(duì)基于人視覺(jué)系統(tǒng)的無(wú)參考JPEG圖像質(zhì)量評(píng)價(jià)方法進(jìn)行了較深入的研究，研究主要內(nèi)容有：
1．在研究當(dāng)前圖像質(zhì)量評(píng)價(jià)方法的基礎(chǔ)上，重點(diǎn)探討了無(wú)參考圖像質(zhì)量評(píng)價(jià)方法，指出了當(dāng)前有影響力的無(wú)參考評(píng)價(jià)方法。
2．將人視覺(jué)系統(tǒng)特征歸納為掩蓋效應(yīng)和視覺(jué)敏感度兩個(gè)特征。針對(duì)掩蓋效應(yīng)，在DCT域分別采用不同的數(shù)學(xué)模型提取紋理邊沿掩蓋效應(yīng)和亮度掩蓋效應(yīng)特征，作為度量掩蓋效應(yīng)特征的指標(biāo)。針對(duì)人視覺(jué)敏感度，則通過(guò)使用不同的濾波算子提取出最能反映人視覺(jué)敏感度的邊沿幅度和長(zhǎng)度、背景活動(dòng)度和亮度這四個(gè)特征。實(shí)驗(yàn)結(jié)果顯示，所采用的數(shù)學(xué)模型提取的特征均具有良好的區(qū)分度。
3．提出一種基于掩蓋效應(yīng)的無(wú)參考圖像質(zhì)量評(píng)價(jià)方法。采用了DCT離散檢測(cè)塊提取出最能反映塊效應(yīng)的掩蓋效應(yīng)值，通過(guò)Minkowski合成法合成一個(gè)可以反映掩蓋效應(yīng)的評(píng)價(jià)指標(biāo)，實(shí)現(xiàn)評(píng)價(jià)圖像質(zhì)量的功能。實(shí)驗(yàn)對(duì)比顯示，該評(píng)價(jià)指標(biāo)能夠較好的捕捉人視覺(jué)注意機(jī)制，較好的反映圖像質(zhì)量的平均主觀評(píng)價(jià)值。
4．提出一種新的基于人視覺(jué)敏感度的無(wú)參考圖像質(zhì)量評(píng)價(jià)方法。采用支持向量回歸神經(jīng)網(wǎng)絡(luò)尋找和逼近圖像質(zhì)量評(píng)價(jià)中人視覺(jué)敏感度特征與平均主觀評(píng)價(jià)值之間的函數(shù)關(guān)系，利用邊沿幅度和長(zhǎng)度、背景活動(dòng)度和亮度等視覺(jué)敏感度特征，實(shí)現(xiàn)符合人視覺(jué)特征的無(wú)參考圖像質(zhì)量評(píng)價(jià)功能。實(shí)驗(yàn)表明，支持向量回歸神經(jīng)網(wǎng)絡(luò)的自主學(xué)習(xí)能力能夠自動(dòng)增添新樣本的特征，具有優(yōu)良的泛化能力和普適性，所得到的圖像評(píng)價(jià)結(jié)果與平均主觀評(píng)價(jià)值有較高的一致性，充分體現(xiàn)了人視覺(jué)特征在圖像質(zhì)量評(píng)價(jià)中的作用。
關(guān)鍵詞 人視覺(jué)敏感度；無(wú)參考；圖像質(zhì)量評(píng)價(jià)；支持向量回歸；神經(jīng)網(wǎng)絡(luò)



Abstract
Image quality assessment had played an important role in the image processing. The no-reference image quality eva luation method had gained a universal attation. Because the essence of image serving people determined that its assessment result must accord with human visual system character, the purpose of image quality assessment is derived to design an effective algorithm which is highly consistent with human sbjective eva luation value of visual perception.The most popular and widely used image format in the Internet as well as in digital cameras happens to be JPEG. Therefore, research object of this paper is NO-reference JPEG image quality assessment based human visual system. The main research results of this paper can be summarized as follows:
1. Summarizing the image quality assessment methods. The no-reference image quality measurement is stressed and then several important no-reference methods are introduced.
2. In this paper, texture edge masking and luminance masking characters are respectively extracted using several of mathematical models and then integrated into a masking map. Human visual sensitivity features such as edge amplitude, edge length, background activity and background luminance are extracted through several filtering operators. The experimental results show that the extracted features all have better discrimination.
3．The No-Reference image quality assessment metric based on masking is presented to predict JPEG image qulity. The masking values on the 8×8 block boundaries are extracted using DCT block discontinuity detection and then can be easily pooled with a Minkowski summation to generate the Mean noticeable Blockiness Score to eva luate image quality. The compared experimental results demonstrate that the metric is highly consistent with mean subjective score, well eva luating image quality.
4. The No-Reference image quality assessment metric based on human visual sensitivity is presented. The support vector regression naural network algorithm is used to search and approximate the functional relationship between human visual sensitivity features and mean subjective score. Then, the measuring of visual quality of JPEG-coded images was realized considering human visual sensitivity features such as edge amplitude, edge length, background activity and background luminance. Experimental results prove that its better generalization performance can add the new features of the sample automatically. Compared with other image quality metrics, the experimental results of the proposed metric exhibit much higher correlation with perception character of HVS. And the role of HVS feature in image quality index is fully reflected.
Keywords human visual sensitivity; NO-Reference; image quality assessment; support vector regression; naural network





















目 錄
摘 要 I
Abstract II
第1章 緒論 1
1.1研究背景與意義 1
1.2研究現(xiàn)狀與發(fā)展趨勢(shì) 2
1.2.1 圖像質(zhì)量評(píng)價(jià)的研究現(xiàn)狀 2
1.2.2 圖像質(zhì)量評(píng)價(jià)的發(fā)展趨勢(shì) 3
1.3 論文的研究?jī)?nèi)容及其主要成果 4
1.4 論文的組織結(jié)構(gòu) 4
第2章 圖像質(zhì)量評(píng)價(jià)綜述 7
2.1 引言 7
2.2 圖像質(zhì)量主觀評(píng)價(jià)方法 7
2.3 圖像質(zhì)量客觀評(píng)價(jià)方法 8
2.3.1 全參考圖像質(zhì)量評(píng)價(jià)方法 8
2.3.1.1 基于誤差統(tǒng)..