Emotional Landscape Image Generation Using Generative Adversarial Networks

Chanjong Park, In-Kwon Lee; Proceedings of the Asian Conference on Computer Vision (ACCV), 2020

We design a deep learning framework that generates landscape images that match an given emotion. We are working on a more challenging approach to generate landscape scenes that do not have main objects making it easier to recognize the emotion. To solve this problem, deep networks based on generative adversarial networks are proposed. A new residual unit called emotional residual unit (ERU) is proposed to better reflect the emotion on training. An affective feature matching loss (AFM-loss) optimized for the emotional image generation is also proposed. This approach produced better images according to the given emotions. To demonstrate performance of the proposed model, a set of experiments including user studies was conducted. The results reveal a higher preference in the new model than the previous ones, demonstrating the production of images suitable for the given emotions. Ablation studies demonstrate that the ERU and AFM-loss enhanced the performance of the model.