Image Morphing Sequences

Abstract

Image morphing is a technique for creating smooth transitions from one image to another and it has applications in the movie industry as well as the medical industry. The simplest kind of morph can be achieved by fading or cross-dissolving between the source and target images. A more convincing morph can be achieved by combining the cross-dissolve with an image warping algorithm. It is a topic which has been extensively researched starting from the 80s and various different methods have been developed for creating convincing looking morphs. Research in this field has a heavy focus on creating high quality and natural looking morphs. One of the main problems of image morphs is ghosting which occurs if the source and target images are not properly aligned or the content of the images is very different. Image warping techniques can be used to reduce ghosting when the images are not aligned by aligning corresponding features. Ghosting typically results in unnatural looking morphs. In this thesis a method is presented to reduce ghosting artifacts in morphs when images do not have corresponding points. The method uses intermediate images to create smaller intermediate morphs. By chaining these morphs together, a more convincing morph can be achieved between the original source and target images. One of the challenges with this method is image selection. Intermediate images require an image data set from where a smaller subset of images is selected for the sequence of image morphs. For optimization, the subset should be as small as possible while producing high quality intermediate morphs. For the experiments in this thesis, a comparison was done between two different methods of image selection. The first method was to select each image manually. In this case the selection was done based on subjective evaluation. The other method was an algorithmic approach. The algorithmic approach takes advantage of a feature based similarity metric, which calculates a similarity score between images. The experiments were conducted with two different data sets of rotating cars. From the first data set a manual selection yielded in 14 images and the algorithmic approach picked 12 images. From the second data set a manual selection yielded in 13 images and the algorithmic approach picked 12 images. The quality of morphs created from images picked by the algorithmic approach were of similar quality to morphs created from manually picked images. The results of the experiments were promising. However, due to the limited scope of this thesis, larger and more in-depth experiments should be conducted.