In this paper, we propose a new de-noising algorithm to handle the mixture of salt and pepper noise and Gaussian noise in videoconference. The algorithm detects the type of noise according to the human visual characteristics first, and then uses median filter to remove salt and pepper noise, and spatial-temporal adaptive filter to remove Gaussian noise. We analyze which parts of the algorithm can be parallelly implemented and implement them using GPU. Experiment results show that our algorithm can remove mixed noise effectively while preserving the edges and details in video, which leads to less bandwidth usage for transmission at the same time. Our implementation with GPU is able to process video frame of resolution 1920*1080 within 25 milliseconds, so it can be used in real-time 1080P/30fps videoconference.