Face detection using DCT coefficients in MPEG video
Jianhua Wang, Mohan Kankanhalli · 2002
We present a neural network-based frontal face detection system, which is completely implemented in the compressed domain. The features used for this purpose are the DCT components of Y, Cr and Cb available from the compressed data of I-frames in MPEG videos. Since DCT coefficients captures frame information concisely, use of DCT features reduces the complexity of the neural network used in the algorithm. In addition, it increases the computational efficiency. The data is used in two stages: in the first stage, a skin color filter, based on Cr and Cb DCT information, is used to locate skin regions. In the second stage, a 4×4 blocks sized window is used to scan the skin regions in the compressed domain image to extract Y-DCT features. A neural network then is trained using these DCT features to classify patterns as faces or non-faces. The preliminary results obtained are encouraging enough to continue research in this direction. 1.