Face-aware Offset Calculation for Video Deblurring

The patent introduces landmark heatmaps as an intermediary element to facilitate accurate alignment. Instead of directly aligning blurry frames, the method first detects landmark heatmaps (representative points such as eyes, nose, mouth) from the blurry target frame and neighboring frames, then uses these heatmaps to calculate offset maps that guide the alignment process. This intermediary approach enables accurate correspondence estimation even when the target frame is severely blurry, resolving the contradiction between maintaining processing speed and improving alignment accuracy.

The patent applies preliminary action by performing landmark heatmap detection and offset map calculation before the main frame alignment and enhancement processes. By pre-computing the offset maps using landmark heatmaps from blurry frames, the method prepares accurate alignment guidance in advance, which then facilitates the subsequent CNN-based enhancement process. This preliminary preparation reduces the difficulty of alignment during the main processing stage, even when target frames are very blurry.

The patent replaces traditional mechanical alignment systems (which rely on direct pixel-level comparison and iterative optimization) with a heatmap-based offset calculation mechanism. Instead of using complex mechanical search algorithms to find correspondences between blurry frames, the method substitutes this with a learned heatmap representation that directly provides offset information. This substitution maintains processing speed while dramatically improving alignment accuracy and thus deblurring performance.