(英) |
In this paper, we investigate deep image compression towards universal usage. In image compression, it is desirable to be able to compress not only natural images but also images in a wide range of domains such as processed photographs, line drawings, and illustrations. However, deep image compression has been generally studied only for natural images and little has been studied for non-natural images. In this study, we first validate the existing deep image compression models using a dataset consisting of multiple domains. Then, we train a compression model on multiple domains and examine the performance on the training domains and unseen domains during training. This method is a baseline method for domain generalization and multi-domain learning. In experiments, we show that deep image compression methods trained on natural images achieve lower performance than traditional methods, especially at higher rates. We also show that while the average performance across multiple domains is higher when training on multiple domains than when training on a single domain, the best performance in each domain is achieved when training on only the evaluation domain. |