| Abstract: |
Abstract—In this paper we present a semantic categorization of images, and discuss the change in the power spectra of these images before and after data hiding. Data hiding alters various natural qualities of the host image, one of which is the image power spectrum. In this paper, we classify a large image database into several categories based on characteristics such as atmospheric details, backgrounds, depths, resolutions, etc. For each category, we calculate the slope of the loglog graph of the spatial frequency versus the amplitude for the marked and unmarked images. We note that in the case of spatial data hiding the average value of the slope for the marked images is at least 45.73% higher than that of the unmarked images. Also in the cases of transform domain data hiding we note that the average value of the slope for the images marked using a discrete cosine (wavelet) transform ( DC(W)T) based technique is higher by at least 3.41% (15.76%). For a commercial data hiding software namely Digimarc corp.’s MyPictureMarc 2005 V1.0, the average value of ® for the marked images is at least 30.61% higher than that of the unmarked images. The maximum percent differences in the value of the slope for all of the four classes of data hiding algorithms are 65.89%, 15.76%, 51.13% and 42.19% for the spatial, DCT, DWT and Digimarc techniques respectively. Based on the observations above we propose to visually classify a test image into one of the semantic categories as in this paper and a slope value deviating from other images in same class will reveal the presence of hidden data in the test image. |
