DCBlock: Efficient module for unpaired image to image translation using GANs

초록

Recently, as various image-to-image translation studies have been progressed, it is possible to generate high-quality images. In particular, generation models using unpaired data produce meaningful results even building data at a low cost. However, these studies, which are based on Generative Adversarial Networks (GANs), is composed a very heavy architecture. Unlike the commonly used other deep learning models, generally the GANs model consists of two or more in a particular case deep architecture, which has a large computational cost. To solve this limitation, this paper proposes an efficient generator module called DCBlock (Depthwise separable Channel Attention Block). DCBlock consists of a depthwise separable convolution with a relatively low computational cost to replace the standard convolution commonly used in the image to image translation, and channel attention to compensate for information loss caused by depthwise separable convolution. DCBlock showed similar performance to the existing original model while reducing the number of parameters that represents the amount of computation by up to 91.6%. Besides, we experiment with the proposed method for various novel researches and prove that the problem is solved. © 2020 Knowledge Systems Institute Graduate School. All rights reserved.

키워드