Skip to main content
— Presentation at ICANN 2017 —
October 18, 2017
Fig. 1 Abstract of proposed method
International Conference on Artificial Neural Networks (ICANN) is the annual conference of European Neural Network Society (ENNS). 26th conference (ICANN 2017) was held in Alghero, Italy, from 11th to 14th September, 2017 and attendees from 33 countries had a discussion on 180 presentations.
The Scopes of ICANN is a deep learning including algorithm, application, modeling of nervous system, and so on. We made a presentation titled "Parallel-pathway Generator for Generative Adversarial Networks to Generate High-Resolution Natural Images", which is an image generation method based on Generative Adversarial Networks (GAN). GAN is one of deep learning methods and it can generate new data from existing dataset. The main problem of previous GAN is that it is difficult to generate a high-resolution image (e.g. upper 128×128) and a rectangular image (e.g. 256×128). In the conference, we reported experimental results of two image generation tasks: high-resolution image generation and rectangular image generation with our proposed methods.
Fig. 2 Network structure
In a previous image generation technique based on GAN, an image generation network had a single pathway deep convolutional neural network structure. On the other hand, we proposed the parallel-pathway structure which utilizes two or more networks with different capacities (Fig. 2). We applied our method to two different image generation tasks: high-resolution human face image with various backgrounds and highway road rectangular image. The generation task of the human face includes different backgrounds, and we show that our network allows for separation of the face and the background. In addition, we then demonstrated that our network with the same parameters and structure of hidden layers could be applied to generate road images with different aspect ratios.
Since the proposed method enables to generate new data as augmentation, we expect that it will be used as data preparation for machine learning when it is difficult to obtain massive data. Further studies would be required for an application field of the proposed method.