Super-Resolution of Speed Signs in Video Sequences

This paper considers the problem of generating a high resolution (super resolved) image of an object that is visible in a sequence of low resolution video frames. We are motivated by applications where the object is moving in the video sequence, due either to movement of the camera or of the object itself. In such cases, accurate sub-pixel image registration can be significantly improved by stochastic filtering for the estimates of inter-frame motion displacement. We use the expectation-maximization (EM) framework to formulate the coupled image registration filtering and super resolution problem. The expectation step is solved as a Bayesian smoothing algorithm based on a motion displacement model correlated to successive frames image registration with the previous super resolved image. The super resolved image is updated as a maximum likelihood estimate based on the new image registration parameters. Results obtained show excellent convergence, robustness and demonstrate quantitative improvement in image quality for real world sequences of speed signs obtained from a moving vehicle.