Real-time dictionary based super-resolution of surveillance video streams and targets

Real-time super-resolution within surveillance video streams is a powerful tool for security and crime prevention allowing for example, events, faces or objects such number-plates and luggage to be more accurately identified on the fly and from a distance. However, many of the state of the art approaches to super-resolution are computationally too expensive to be suitable for real-time applications within a surveillance context. We consider one particular contemporary method based on sparse coding,1 and show how, by relaxing some model constraints, it can be sped up significantly compared to the reference implementation, and thus approach real-time performance with visually indistinct reduction in fidelity. The final computation is three orders of magnitude faster than the reference implementation. The quality of the output is maintained: PSNR of the super-resolved images compared to ground truth is not significantly different to the reference implementation, while maintaining a noticeable improvement over baseline bicubic-interpolation approach.