Improving the Performance of MPEG Compatible Encoding at Low Bit Rates Using Adaptive Neural Networks

Abstract

A new approach is presented in this paper for improving the performance of MPEG encoders, especially in videophone or videoconferencing applications, through allocation of a greater number of bits in objects that belong to the foreground of image frames, than in objects that belong to the background. A human face and body detector followed by a neural network classifier are used for foreground/background object extraction. The derived image segmentation is used to modify the rate control of MPEG schemes so as to allocate more bits to foreground objects than to background, while retaining compatibility with MPEG encoders. Experimental results are presented, including image sequences with complex backgrounds, which illustrate the performance of the proposed scheme. Both a subjective image quality improvement and a PSNR increase of about 1.35 db on average have been obtained.