In this paper, we propose a CNN-feature-based image retrieval method to find the ideal style image that better translates the material of an object. An ideal style image must share semantic information with the content image, while containing distinctive characteristics of the desired material. Therefore, we first refine the search by selecting the most discriminative images from the target material. Subsequently, our search process focuses on the object semantics by removing the style information using instance normalization whitening. Thus, the search is performed using the normalized CNN features. In order to translate materials to object regions, we combine semantic segmentation with neural style transfer. We segment objects from the content image by using a weakly supervised segmentation method, and transfer the material of the retrieved style image to the segmented areas. We demonstrate quantitatively and qualitatively that by using ideal style images, the results of the conventional neural style transfer are significantly improved, overcoming state-of-the-art approaches, such as WCT, MUNIT, and StarGAN.

In this paper, we tackle a challenging domain conversion task between photo and icon images. Although icons often originate from real object images (i.e., photographs), severe abstractions and simplifications are applied to generate icon images by professional graphic designers. Moreover, there is no one-to-one correspondence between the two domains, for this reason we cannot use it as the ground-truth for learning a direct conversion function. Since generative adversarial networks (GAN) can undertake the problem of domain conversion without any correspondence, we test CycleGAN and UNIT to generate icons from objects segmented from photo images. Our experiments with several image datasets prove that CycleGAN learns sufficient abstraction and simplification ability to generate icon-like images.

We propose a texture synthesis method based on generative adversarial networks, focusing on a cultural emblem, called Batik, of southeastern Asian countries. We propose a two-stage training approach to construct the network, first generating patches and then combining patches to generate the entire Batik image. Regular repetition and synthesis artifact removal are jointly considered to guide model training. In the evaluation, we show that the proposed generator fuses two Batik styles, removes blocking artifacts, and generates harmonious Batik images. Qualitative and quantitative evaluations are provided to show promising performance from several perspectives.

People change their hairstyles to make their appearance attractive, however it is difficult to determine which hairstyles are attractive. In this study, we aim to recommend a hairstyle that improves the attractiveness for an input face using attractiveness evaluation and image generation by deep learning. In the experiment, we first learned the attractiveness and obtained results similar to human intuition. Second, the hairstyle of the input image was changed using two methods: hairstyle attribute conversion and face swapping. Finally, a comparison experiment was performed by subjectively evaluating the input image and the image obtained by the proposed method. As a result, the proposed method was able to generate images with high evaluation value.

YouTubers have recently become highly popular. Generating eye-catching thumbnails is an important factor in attracting viewers. In this study, we propose an automatic YouTube-video-thumbnail generation method that ensures the following: rich facial expression of the YouTuber in the frame, clear presentation of the subject of the video, and clear description of the content through a headline. We compared thumbnails generated by the proposed method with those generated by existing methods or those of actually posted videos (i.e., ground truth), and we evaluated the results.