For adaptive perceptual quantization in video coding, the image sequence is segmented into macroblocks and each macroblock is classified based on its local perceptual sensitivity. As the size of the macroblock increases, however, it is hard to classify each macroblock into a certain class. This paper describes a new adaptive perceptual quantization algorithm by the macroblock classification based on the neural network model. The neural network model is well known as an efficient classifier and has been widely used as a pattern classifier. In order to construct the neural network model as a classifier, the most important thing is to determine proper target classes of macroblocks. The target class of each training macroblock is determined by the two steps. First, a simple classification algorithm based on human visual perception is performed to obtain the preliminary target classes for training macroblocks. Second, the obtained target classes are refined by visual quality test of the coded sequence. The neural network model is then trained using the refined training set. Since the macroblocks with classification error are reassigned to more proper classes, the proposed classifier can reduce the classification error.
The performance of the proposed neural network based classifier is investigated by computer simulation. In comparison with the non-adaptive quantization scheme and the adaptive quantization scheme in the MPEG-2 TM5, the proposed scheme is proven to provide improved perceptual quality in the video coding.
