High-performance image deinterlacing using optical flow and artifact post-processing on GPU/CPU for surveillance and reconnaissance tasks

The necessity to process interlaced images in surveillance, reconnaissance, or further computer vision areas should be a topic of the past. But, for different reasons, it is not. So, there are situations in practice, in which interlaced images have to be processed. Since a lot of algorithms strongly degrade when working with such images directly, a usual method is to double or interpolate image lines in order to discard one of the two enclosed image frames. This is efficient but leads to weak results, in which half of the original information is lost. Alternatively, a lot of valuable computation time has to be spent to solve the highly complex motion compensation task in order to improve the results significantly. In this paper, an efficient algorithm is presented to solve this dilemma. First, the algorithm solves the complex 2-D mapping problem using the best state-of-the art optical flow method that could be found for this purpose. But, of course, for different physical reasons there are regions which cannot properly be handled by optical flow by itself. Therefore, an efficient post-processing method detects and removes remaining artifacts afterwards, which is the main contribution of this paper. This method is based on color interpolation incorporating local image structure. The presented results document the overall performance of the approach with respect to obtained image quality and calculation time. The method is easy to implement and offers a valuable pre-processing for a lot of computer vision tasks.