Statistical Separation Of Compressed And Uncompressed Natural Color Images
53.526, Tuesday, May 14, 8:30 am - 12:30 pm, Vista Ballroom
Michele Saad1, Alan Bovik1, Lawrence Cormack1; 1The University of Texas at Austin
Statistical models of luminance images have been applied to a wide variety of computational visual processing tasks such as quality assessment, image restoration, and perceptual image repair. However, the statistics of chromatic images have been much less extensively studied. We have studied the distribution of the HSV hue channel in natural images. Using a corpus of natural color images, we find a model of the joint hue distribution of neighboring image samples. Using this model, we compare the statistics of natural scenes both to a baseline (spatially random sample pairs) and to the statistics of images distorted by JPEG compression. We show that there is a distinct difference in the hue statistics between the three groups. We then repeat the analysis on the L*a*b* space chroma channels, and find that the difference between the groups is less pronounced in L*a*b* space than in HSV space. By randomly selecting 80% of the natural color images in the database, we formed a dictionary of "words" representative of naturalistic color images. Each word was expressed as the joint histogram of the neighboring pixel hue values. From this dictionary, we obtained a coarser representation of the naturalistic color image signatures by a process of vector quantization. We show that the pristine images in the remaining 20% of the image corpus lie much closer in Euclidean distance to the cluster centers than do the distorted versions (and the separation is highly significant). These results are interesting in their own right, and can also be useful for improving upon computational visual processing tasks such as quality assessment, whereby the larger the distance between a distorted image and the natural color scene dictionary, the larger the expected perceived distortion.