Merging Facial Images With Visual Feature Extraction

Research on computer vision has been booming ever since it became active in the 1970s. More rapid advances in this area were seen in 1990s, driven by the always-decreasing price/performance ratio of computing coupled with the recent drop in digital image acquisition cost. Among the many topics emerged in this literature, face detection and recognition has received significant attention. Automatic facial feature extraction and localization in images now can be done with acceptable error rate.On the other hand, morphing is an image processing technique popularized in 1980s for visual effects in movies and television. It enables fluid transformation from one image into another. These methods require features of the two images being mapped accurately to eliminate the undesired overlapping effect from cross-dissolving image pixels. Normally this means tedious work that has to be done manually.It seems natural to achieve automatic morphing by incorporating computer vision techniques, because the image features required by morphing are the very output of some vision algorithms. In this thesis, we present a framework to automatically create an in-between image from two or more facial images.First, we propose a coarse-to-fine procedure to extract facial features for latter processing. Existence of face as well as the approximate location is determined by face detection, and utilized for estimating pose parameters of facial features. Then accurate shape and pose parameters and be obtained, i.e. facial features can be extracted, by face alignment.Second, we present the algorithm for in-between image creation, which is similar to polymorph. The mapping of correspondent pixels can be established from their relative positions against the feature locations, thus the whole process is fully automatic. We also describe in this thesis a preprocessing step that reduces color difference of facial area to improve the result from images of significantly different illumination conditions, and a post-processing step to further enhance the final result.