Department of Computer Engineering

Creating a high dynamic range (HDR) image from differently exposed images is a complex problem, especially for dynamic scenes. Most existing algorithms often match the luminance of frames to a reference image and merge exposure-matched frames into a single HDR image. However, due to the motion in the exposure bracket, this process produces artifacts such as ghosting or tearing in the merged image. These issues not only degrade the perceptual quality of the final result but also affect the accuracy of face recognition algorithms. In this thesis, we propose a face-aware optical flow-based deghosting algorithm for HDR image generation using multiple images with different exposures. We focus on preserving the facial features across the exposure bracket and in the merged HDR results. The proposed approach alters a previous learning-based optical flow method for image alignment before merging, significantly improving facial recognition accuracy while maintaining real-time performance.