图像质量评价:融合视觉特性与结构相似性指标

图像质量评价是多媒体领域的一个基本问题.在现有结构相似性方案的基础上,增加最小可辨失真和视觉显著性评价用于构造一个新的图像质量评价指标.首先,考虑到人眼视觉系统的掩蔽效应,设计了失真图像的修正模型,以Sigmoid函数为基础,利用最小可辨失真阈值和图像的像素变化绝对值构建失真图像修正模型,对失真图像进行修正,使其符合人眼的感知效果;然后,考虑到人眼视觉系统的视觉注意特性,设计了图像区域权重模型,利用视觉显著性图像表征图像感兴趣内容,作为图像不同区域的权重;最后,计算修正后的失真图像与原图像的局部结构相似性,并利用区域权重对局部区域质量进行加权平均获得全局图像的质量评价值.实验结果表明,与同类度量指标相比,本文提出的指标在评价图像局部质量方面,更符合人的主观感知效果;在主观数据拟合方面,其均方根误差、相关系数等指标都得到提升;在运行效率方面,具有一个适中的计算复杂度,远低于性能优越的度量指标的运行时间.

Image quality assessment: Combining the characteristics of HVS and structural similarity index

Image quality assessment is a basic problem in the multimedia field. A new image quality metric is constructed based on Structural Similarity (SSIM) index, by exploiting Human Visual System (HVS) characteristics. First, considering the masking effects of HVS, the distorted image is preprocessed by a distortion model with the input of the error between it and the original one and just noticeable distortions (JNDs) derived from a human visual model, where Sigmoid function is explored. The process makes the visible errors in the modified distorted image to be more notable. Second, considering the visual attention characteristics of HVS, the image area weight model is designed to quantify the importance of image local regions for the visual quality. The interesting content of an image can be represented by the saliency image, from which the weights of different regions are obtained. Finally, the local SSIM between the modified distorted image and original image is calculated, and the global image quality metric can be expressed by weighting all local quality with the normalized regional weights. Compared with the state-of-the-art image metrics, the proposed metric fits subjective visual quality better in evaluating the local image quality, has better performance in terms of the mean square error, correlation coefficient and other indicators for predicting the subjective image quality, and has a moderate computational complexity, well below the run time of the superior performance metrics.