An Improved Statistical Model of Appearance under Partial Occlusion

The Appearance Models (AMs) are widely used in many applications related to face recognition, expression analysis and computer vision. Despite its popularity, the AMs are not much more accurate due to partial occlusion. Therefore, the authors have developed Robust Normalization Inverse Compositional Image Alignment (RNICIA) algorithm to solve partial occlusion problem. However, the RNICIA algorithm is not efficient due to high complexity and un-effective due to poor selection of Robust Error Function and scale parameter that depends on a particular training dataset. In this paper, an Improved Statistical Model of Appearance (ISMA) method is proposed by integration techniques of perceptual-oriented uniform Color Appearance Model (CAM) and Jensen-Shannon Divergence (JSD) to overcome these limitations. To reduce iteration steps which decrease computational complexity, the distribution of probability of each occluded and un-occluded image regions is measured. The ISMA method is tested by using convergence measure on 600 facial images by varying degree of occlusion from 10% to 50%. The experimental results indicate that the ISMA method is achieved more than 95% convergence compared to RNICIA algorithm thus the performance of appearance models have significantly improved in terms of partial occlusion.