Title:High Frequency Content based Stimulus for Perceptual Sharpness Assessment in Natural Images

Abstract:A blind approach to evaluate the perceptual sharpness present in a natural image is proposed. Though the literature demonstrates a set of variegated visual cues to detect or evaluate the absence or presence of sharpness, we emphasize in the current work that high frequency content and local standard deviation can form strong features to compute perceived sharpness in any natural image, and can be considered an able alternative for the existing cues. Unsharp areas in a natural image happen to exhibit uniform intensity or lack of sharp changes between regions. Sharp region transitions in an image are caused by the presence of spatial high frequency content. Therefore, in the proposed approach, we hypothesize that using the high frequency content as the principal stimulus, the perceived sharpness can be quantified in an image. When an image is convolved with a high pass filter, higher values at any pixel location signify the presence of high frequency content at those locations. Considering these values as the stimulus, the exponent of the stimulus is weighted by local standard deviation to impart the contribution of the local contrast within the formation of the sharpness map. The sharpness map highlights the relatively sharper regions in the image and is used to calculate the perceived sharpness score of the image. The advantages of the proposed method lie in its use of simple visual cues of high frequency content and local contrast to arrive at the perceptual score, and requiring no training with the images. The promise of the proposed method is demonstrated by its ability to compute perceived sharpness for within image and across image sharpness changes and for blind evaluation of perceptual degradation resulting due to presence of blur. Experiments conducted on several databases demonstrate improved performance of the proposed method over that of the state-of-the-art techniques.