Hybrid video-frame pre-processing architecture for HD-video

As imaging devices begin to support full frame rate video at 720p and higher resolutions image processing algorithms which operate on full image frames become severely challenged due to both the processing costs of operating on full-frame resolution images and the bandwidth challenges of moving image data to and from system memory. In this paper we outline a novel hybrid hardware architecture which enables a range of image processing primitives to be directly generated from a video frame as it passes through the standard imaging pipeline. These primitives can be either efficiently stored with a sequence of video frames, or depending on the computational capabilities of the device they may be recombined with the original video frame data, concurrently with the acquisition of the following video frame, to implement a range of image processing functions prior to MPEG video data compression. Examples include color segmentation, face tracking, frame-to-frame registration, motion stabilization, skin detection, dynamic range enhancement, global and local histograms and computation of integral image functions.