D. Zhong and I. Defée (Finland)
DCT Transform, Quantization, Feature extraction, Facial feature, Compressed-domain
Images and video are currently predominantly handled in compressed form. Block-based compression standards are by far the most widespread. It is thus important to devise information processing methods operating directly in compressed domain. In this paper we investigate this possibility on the example of simple facial feature extraction method based on the Discrete Cosine Trans form (DCT) coefficients. According to our experiments, most horizontal information of face images is mainly distributed over some key features. After applying block transform and quantization to the face images, such significant information become compact and obvious. Therefore, by evaluating the energy of the specific coefficients which are representing the horizontal information, we can locate the key features on the face. The approach is tested on FERET data-base of face images and good results is provided despite its simplicity.
Important Links:
Go Back
