Research On Face Detection In The DCT Compressed Domain

As one of the most representative characteristics of individuals, human face iscrucial for image/video presentation. Similarity can be determined among differenthuman faces. However how to describe and utilize these common features is one ofthe key issues. With the reality that most of image/video are stored and transmittedin compressed format, it is promising to investigate how human face will be detectedin compressed domain. The major research effort of this thesis is to study how toextract facial features and detect human faces in compressed domain. This thesis presents a multimodal Gaussian model based approach for facedetection in compressed domain, which combines neural network classifier and skintone verification module. The approach utilizes multimodal Gaussian model toapproximate the face patterns distribution in the feature space. The face samples aregrouped into several clusters by modified k-means algorithm. Meanwhile, theclusters of non-face samples are applied to refine the boundaries of the distributionof face patterns. Distribution based face model is composed of the clusters of faceand non-face patterns. Based on the face model, certain measurement is designed. Asa classifier, the BP neural network implements the identification. In addition, for thecolor image, the information of skin tone is fully utilized, and a skin tone verificationmodule is designed. This thesis also discusses some problems relevant to thecompressed domain, including, blocking effect resulted from blocking operation inthe compression standard, the derivation of merging DCT coefficients by basisfunctions transformation, and scaling image with fractional factor directly incompressed domain. Human face detection in compressed domain has versatile application. This thesishas a fundamental research on the subject of human face detection in compresseddomain, and some decent results have been obtained. It will be a promising researchdirection for multimedia content analysis to have traditional pattern detected, in thecompressed domain.