AbstractImage generation is one of the most important fields in Generative Adversarial Networks (GANs), as well as one of the most challenging ones, because of the challenges of image manipulation and reconstruction.
Therefore, in this project, we will focus on the generation of images of cats when the facial features (e.g. Eyes, Ears, and Nose) of a cat are modified, such that we could make the eye larger, or make it in a slightly different position, or move the nose or ear to a different position. This will be done by representing the facial features of a cat using elliptical shapes fitted on each of these facial features, then these representations will be used as an input for pix2pix to reconstruct the faces of cats.
Furthermore, the dataset used in this project is CAT dataset, which contains 10 000 images of cats with annotations. These annotations are (x, y) positions of the facial features of the cats. Moreover, the elliptical shapes which will represent the facial features of the cats, are fitted according to annotations supplied with the dataset. In addition, some of the segmentation are modified to check if the model is useful for image editing.
Finally, the results show that our approach is successful in representing the facial features of cats, compared to the approach conducted by Dekel et al. , which is based on contours. In addition, these representations of the facial features of cats provided pix2pix with high level of information, this was observed from the results of the modified segmentation images.
|Date of Award
|20 Nov 2019
- Computer Vision
- Image Editing
- Deep Neural Network