基于高通差异性特征的图像质量评估方法

High-Pass Difference Features Based Image Quality Assessment

现有的图像质量评估只能判断单一失真方式下失真图像的质量优劣。为了改进这一缺点,根据自然场景统计信息的图像特征,提出基于高通滤波下RGB差异性的图像质量评估方法,通过局部归一化亮度,提取RGB通道差异性、图像梯度、图像锐度、及图像对比度等特征,利用逻辑回归训练,最终得到无参考图像质量评估模型。实验结果表明,方法对各类型失真图像质量评估准确率较高,特别对多种失真类型混合的测试集时,具有明显优势。

Current methods of image quality assessment only can assess the quality of images under the same type of image distortion. In order to fix such weaknesses, this paper is designed based on the image features of natural scene statistics and proposes a new metric method using high-pass filter for detecting features. The approach computes locally the normalized luminance;selects features such as the difference of RGB channels via high-pass filter, image gradient, sharpness, contrast, etc.;and analyzes and gathers features in the metric method trained by logistic regression. Experimental results show that the proposed method can work efficiently under multiple distortion types and is significantly better than current no-reference image quality assessment methods under the test sets, which gather multiple distortion types.