Image Quality Assessment Using Level-of-Detail
Image quality assessment is a very challenging problem in image processing. Most image quality metrics are currently based on super cial variations in pixel values, perceptual models and structural changes. As well, most are full-reference metrics in which a corrupted image is compared with an \original" or \perfect" version of that image. In many practical settings, however, this \perfect" image is not available. At this time, no metric is able to genuinely replicate human perception of quality. This research therefore introduces a new image quality model, one that centers on level-of-detail. Furthermore, the proposed techniques operate without any \original" image, making them ideal for real-world applications such as television monitoring. In this research, level-of-detail is employed in both the initial detection and subsequent psychological evaluation of noise, blurring and compression. A given image is tested for noise and blurring using a new measure of detail in the frequency domain. For perhaps the rst time, it enables a machine to separate noisy, blurred and otherwise uncorrupted images. This decision is based on the slope of the cumulative histogram of the spectral energy values. Moreover, noise and blur are appropriately organized as \opposite" phenomena in this model, with noise corresponding to high levels of detail and blur being paired with lesser amounts. Once detected, the precise magnitude and speci c type of noise, whether it be random, Gaussian or salt-and-pepper, or blur, either isotropic or motion, is determined using individual noise and blur metrics. Using an alternate notion of level-of-detail based on the prominent mean shift segmentation algorithm, this time in the spatial domain, blocking and ringing metrics to detect and gauge the e ects of JPEG and JPEG2000 compression errors are de ned. All errors uncovered in an image are psychologically weighed in relation to surrounding image content. The goal is to determine which errors are perceptible and which are masked by neighboring pixels. Further motivating this research are the ndings of an investigation of a popular full-reference metric, namely the structural similarity index. Experimental test cases, a numerical analysis and a theoretical study link this method to the conventional mean squared error.