-
为了提出性能优异的图像质量评价(IQA)模型, 本文基于人类视觉感知特性和图像的灰度梯度、局部对比度和清晰度特征, 提出了一种基于图像内容对比感知的IQA方法. 在该方法中, 首先结合视觉感知特性, 基于物理学中对比度定义, 提出一种图像质量定义及其值计算方法; 之后, 基于灰度梯度共生矩阵, 提出一种图像灰度梯度熵的概念及其值的计算方法, 并基于图像灰度梯度熵、局部对比度和清晰度, 提出一种图像内容及其视觉感知的描述方法; 最后, 基于图像内容特征和图像质量定义, 综合分析, 提出IQA方法及其数学模型. 并且采用5个开源图像数据库中的119幅参考图像和6395幅失真图像对其进行了仿真测试, 同时分析和探讨了52种失真类型对IQA的影响; 另外, 为了说明所提IQA模型的优势, 将其与现有的7个典型IQA模型, 从精度、复杂性和泛化性能上进行了对比分析. 实验结果表明, 所提IQA模型的精度PLCC值在5个数据库中最低可以实现0.8616, 最高可达到0.9622, 其性能综合效益优于7个现有IQA模型. 研究结果表明, 所提IQA方法是有效的、可行的, 所提IQA模型是一个性能优异的IQA模型.Image quality assessment (IQA) plays a very important role in acquiring, storing, transmitting and processing image and video. Using the characteristics of human visual perception and the features of the gray, gradient, local contrast, and blurring of image, an IQA method based on the image content contrast perception is proposed in the paper, which is called MPCC. In the proposed method, firstly, combining with the characteristics of human visual perception, based on the definition of the contrast in physics, a novel definition for image quality and its calculation method are proposed. Then, based on the gray gradient co-occurrence matrix, a novel concept, namely the gray-gradient entropy of image, and its calculation method, are proposed. And based on the gray-gradient entropy, local contrast and blurring of image, a method of describing the image content and their visual perception are proposed. Finally, based on the image content features and the image quality definition, an IQA method and its mathematical model are proposed by comprehensive analysis. Further, the proposed IQA model MPCC is tested by using 119 reference images and 6395 distorted images from the five open image databases (LIVE, CSIQ, TID2008, TID2013 and IVC). Moreover, the influences of the 52 distortion types on IQA are analyzed. In addition, in order to illustrate the advantages of the MPCC model, it is compared with the seven existing typical IQA models in terms of the accuracy, complexity and generalization performance of model. The experimental results show that the accuracy PLCC of the MPCC model can achieve 0.8616 at lowest and 0.9622 at most in the five databases; among the 52 distortion types, the two distortion types, namely the change of color saturation and the local block-wise distortions of different intensity, have the greatest influence on IQA, and the accuracy PLCC values of the seven existing IQA models are almost all below 0.6, but the PLCC of the MPCC model can reach more than 0.68; and the comprehensive benefit of the performance of the MPCC model is better than those of the seven existing IQA models. These results of test and comparison above show that the proposed IQA method is effective and feasible, and the corresponding model has an excellent performance.
[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] -
数据库 LIVE(779) CSIQ(866) TID2008(1700) TID2013(3000) 加权 PLCC 0.9622 0.9586 0.8778 0.8616 0.8915 SROCC 0.9660 0.9569 0.8831 0.8452 0.8854 RMSE 7.4397 0.0747 0.6427 0.6293 — OR 0.1531 0.2690 0.1287 0.1198 — 
下载:
导出CSV
数据库 参数 PSNR VSNR SSIM FSIMc VSI GMSD MAD MPCC CSIQ PLCC 0.8000 0.8002 0.8613 0.9192 0.9279 0.9541 0.9502 0.9587 SROCC 0.8058 0.8106 0.8756 0.9310 0.9423 0.9570 0.9466 0.9569 RMSE 0.1575 0.1575 0.1334 0.1034 0.0979 0.0786 0.0818 0.0748 OR 0.4220 0.3832 0.3535 0.3041 0.2873 0.2742 0.2829 0.2738 LIVE PLCC 0.8723 0.9231 0.9449 0.9613 0.9482 0.9603 0.9675 0.9620 SROCC 0.8756 0.9274 0.9479 0.9645 0.9524 0.9603 0.9669 0.9660 RMSE 13.3597 10.5059 8.9455 7.5296 8.6816 7.6214 6.9073 7.4598 OR 0.2179 0.2151 0.1865 0.1627 0.1853 0.1643 0.1529 0.1606 TID2013 PLCC 0.7062 0.7402 0.7895 0.8769 0.9000 0.8553 0.8267 0.8648 SROCC 0.6917 0.7316 0.7417 0.8510 0.8965 0.8044 0.7807 0.8452 RMSE 0.8887 0.8392 0.7608 0.5959 0.5404 0.6423 0.6975 0.6224 OR 0.1636 0.1552 0.1427 0.1132 0.1045 0.1242 0.1323 0.1179 下载:
导出CSV
失真类别 PSNR VSNR SSIM FSIMc VSI GMSD MAD MPCC 1 Additive Gaussian noise(AGN) 0.9552 0.8319 0.8685 0.9152 0.9527 0.9503 0.8897 0.8706 2 Noise in color comp. (NCC) 0.9256 0.7814 0.8050 0.8873 0.9172 0.9118 0.8438 0.8324 3 Spatially correl. noise (SCN) 0.9525 0.8105 0.8621 0.8989 0.9472 0.9391 0.9008 0.7457 4 Masked noise (MN) 0.8707 0.7715 0.8219 0.8492 0.8203 0.7547 0.8009 0.6943 5 High frequency noise (HFN) 0.9731 0.9061 0.9081 0.9475 0.9655 0.9567 0.9233 0.9090 6 Impulse noise (IN) 0.8887 0.7442 0.7415 0.8171 0.8635 0.7572 0.3206 0.7408 7 Quantization noise (QN) 0.8880 0.8384 0.8702 0.8794 0.8747 0.9110 0.8571 0.8122 8 Gaussian blur (GB) 0.9169 0.9437 0.9634 0.9544 0.9551 0.9099 0.9357 0.9252 9 Image denoising (ID) 0.9640 0.9463 0.9589 0.9652 0.9707 0.9759 0.9645 0.9594 10 JPEG compression (JPEG) 0.9167 0.9386 0.9551 0.9754 0.9858 0.9843 0.9638 0.9509 11 JPEG2000 compression (JPEG2 K) 0.9170 0.9513 0.9658 0.9754 0.9845 0.9812 0.9740 0.9452 12 JPEG transm. errors (JPEG trans.) 0.8104 0.8597 0.9181 0.9176 0.9457 0.9079 0.9001 0.8805 13 JPEG2000 transm. errors (JPEG2K trans) 0.9002 0.8435 0.8801 0.8929 0.9192 0.9085 0.8838 0.8699 14 Non ecc. patt. noise (NEPN) 0.6746 0.6774 0.7773 0.8068 0.8162 0.8133 0.8608 0.8132 15 Local block-wise dist. (LBWD) 0.2410 0.3632 0.6022 0.5542 0.4984 0.6520 0.4187 0.6845 16 Mean shift (MS) 0.8056 0.5160 0.8019 0.7869 0.8021 0.7707 0.6934 0.7720 17 Contrast change (CC) 0.5811 0.4251 0.6026 0.7266 0.6974 0.7111 0.3199 0.8108 18 Change of color saturation (CSS) 0.3294 0.4184 0.4590 0.8228 0.8052 0.4234 0.2846 0.7583 19 Multipl. Gauss. noise (MGN) 0.9204 0.7730 0.7896 0.8660 0.9136 0.8911 0.8529 0.8759 20 Comfort noise (CN) 0.8702 0.9016 0.9022 0.9463 0.9546 0.9562 0.9444 0.8476 21 Lossy compr. of noisy (LCN) 0.9429 0.8960 0.9174 0.9564 0.9636 0.9703 0.9562 0.7889 22 Image color quant. w. dither (CQWD) 0.9308 0.8773 0.8619 0.8911 0.8963 0.9192 0.8779 0.8721 23 Chromatic aberrations (CA) 0.9556 0.9592 0.9770 0.9794 0.9748 0.9737 0.9696 0.9473 24 Sparse sampl. and reconstr. (SSR) 0.9296 0.9477 0.9667 0.9776 0.9808 0.9849 0.9766 0.9349 Max 0.9731 0.9592 0.9770 0.9794 0.9858 0.9849 0.9766 0.9594 Min 0.2410 0.3632 0.4590 0.5542 0.4984 0.4234 0.2846 0.6845 波动范围宽度 0.7321 0.5959 0.5181 0.4252 0.4873 0.5614 0.6920 0.2750 所有整体精度 0.7062 0.7402 0.7895 0.8769 0.9000 0.8553 0.8267 0.8648 下载:
导出CSV
-
[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31]
下载:
计量
- 文章访问数:8917
- PDF下载量:154
- 被引次数:0