Evaluating color and texture features for forgery localization from illuminant maps

TL;DR: Experiments show that the Local Phase Quantization (LPQ) descriptor performs best in identifying the spliced image region from the illuminant map, and different histogram similarity measures including heuristic histogram distance measures, non-parametric test statistics, information theoretic divergences, and cross-bin measures.

Abstract: Images are widely accepted as a record of events even when images are prone to easy manipulations. It is difficult to identify image alterations by the human visual system. Once an image is identified as forged, the next step is to locate forged regions. Recently, distribution of scene illumination across an image has been analyzed to detect forged images and to locate forged image regions. In this paper, we investigate the problem of locating spliced image region based on illumination inconsistency. We investigated the discriminative power of a number of color and texture descriptors in locating spliced image regions. During digital crime investigations, often it is required to detect the spliced face in a group photo. Here, we have selected forged images containing human facial regions where the regions to be compared are of similar object material, human skin regions. We evaluated various color, texture, and combined color-texture descriptors in an unsupervised manner by comparing the distance between the feature vectors to identify the inconsistent image region. We also investigated the performance of different histogram similarity measures including heuristic histogram distance measures, non-parametric test statistics, information theoretic divergences, and cross-bin measures. Experiments show that the Local Phase Quantization (LPQ) descriptor performs best in identifying the spliced image region from the illuminant map.