学习相位一致特征的无参考图像质量评价

No-reference Image Quality Assessment with Learning Phase Congruency Feature

为了有效地评价不同失真类型的图像质量,该文提出一种利用广义回归神经网络(GRNN)学习图像相位一致特征的无参考评价方法。该方法首先使用相位一致模型产生相位一致图像、相位一致协方差最大、最小图像,然后利用灰度-梯度共生矩阵模型计算该3幅图像的梯度熵、原始图像的梯度均值和梯度熵,再加上该3幅图像的均值,共8个特征输入GRNN模型训练学习,预测得到图像质量评价得分。实验结果表明新方法与主观得分有较好的一致性,同时具有可靠的推广性。

In order to assess multi distorted types of image quality effectively, a new no reference image quality assessment method is proposed, which uses General Regression Neural Network （GRNN） model to predict image quality score by learning phase congruency feature. In this method, three images, namely the Phase Congruency （PC） image, the maximum moment of PC covariance and the minimum moment of PC covariance image, are produced by the phase congruency model. Secondly the gradient entropy of the three reproduced images, the gradient mean value and the gradient entropy of the original image are computed by gray-level gradient co-occurrence matrix model, and the mean value of the three images are also calculated. At last all above eight features are fed to GRNN to learn, and predict image quality score. Experimental results demonstrate our algorithm is more consistent with human subjective scores and moreover has credible generalization.