刘玉涛 赵德斌
Abstract: Images are vulnerable to different kinds of distortions, such as blur, noise, blocking, etc., which all degrade the image quality. Among the distorted images, out-of-focus blurred images are frequently-encountered and occupy a large proportion. However, few efforts have been done to quality evaluation for these images. In this paper, a dedicated quality evaluation scheme is devised to automatically infer the quality of out-of-focus blurred images, which is named GPSQ (Gradient magnitude and Phase congruency-based and Saliency-guided Quality model). In GPSQ, a pair of low-level features, including Gradient Magnitude (GM) and Phase Congruency (PC), are extracted to characterize the image local blurriness. Then saliency detection is performed on the image to generate a corresponding saliency map. Finally, the local structure map with the saliency map is weighted to estimate the visual quality of the out-of-focus blurred image. Experimental results demonstrate the proposed GPSQ delivers high consistency with subjective evaluation results.
引言
隨着手机相机的普及使得图像的获取变得非常便捷。然而,不适当的相机参数或者人为拍摄的随意性使得获取的照片很容易遭受失真的干扰。这些失真主要有模糊、噪声、对比度改变等等。在失真图像里面,失焦模糊图像占据很大比例,因此,本文提出一种专门用于失焦图像的质量评价的方法。该方法不仅可以在拍摄过程中检测失焦模糊图像的质量,还可以帮助在拍摄后期剔除失焦模糊的图像。
一般地,图像质量评价方法可分为3类,即:全参考、半参考和无参考评价方法。其中,全参考质量评价方法是指在原图或者无失真图已知的情况下,通过比对原图和对应的失真图来预测失真图的质量。代表性的方法主要有PSNR(peak signal-to noise ratio)和SSIM (structural similarity index) [1]等。具体来说,PSNR的做法是计算均方误差(Mean Squared Error, MSE)来估计图像的质量,MSE越小表示失真图像与原图越接近,从而失真图像的质量越高,相反失真图像的质量越低。SSIM通过计算原图与失真图像之间的亮度、对比度和结构相似性来估计图像的质量。因为有原图可以参考,分析可知全参考评价方法的预测性能是最高的,但是要求原图已知的条件太高,现实应用中往往很难达到。因此,研究者考虑只提取部分原图的信息或特征,通过比较特征来对图像的质量进行预测,这就是半参考的图像质量评价方法。该类方法不需要参考原图全部的信息,只需要提取原图的一些特征,运行时极大地降低了数据量。代表性方法主要有RRED(Reduced Refer