In this work we propose a fast algorithm for retinex computation of a video sequence. Using traditional multi-scale retinex computation one may process each frame of the video sequence individually. But this requires large computation time as it involves convolution of individual frames with large Gaussian masks. To reduce the computation time, we have performed convolution over a single frame and used these values for the subsequent frames. We have observed the results are similar with those obtained by the individual frame processing. However, the results greatly deteriorate when the scene-changes occur in the video sequence. Hence we have also detected the scene changes and made the process adaptive.  download full paper