Segmentation based coding of human face images for retrieval

Abstract

A coding method for face images by segmenting face regions Such as eyes, mouth, etc. is introduced. Each region is removed from the face image, leaving a smoothly interpolated background which is suitable for efficient compression. Extracted region images are coded using vector quantization. The overall method is found to achieve comparable compression ratios with compatible image coding methods. Furthermore, separate encoding of face segments enables us to achieve speed improvements for a face image retrieval system. Vector quantization of semantically meaningful face regions is observed to supply indexes that correspond to information about face similarities