Image Purification through Controllable Neural Style Transfer
Tongtong Zhao, Yuxiao Yan, Ibrahim Shehi Shehu, HaoHui Wei, Xianping Fu · 2018
Recently, progress in learning-by-synthesis proposed training models on synthetic images that can effectively reduce the cost of human and material resources. However, learning from synthetic images still cannot achieve the desired performance due to the different distribution of synthetic images compared to naturalistic images. Naturalistic images are composed of multiform light distribution, which is a characteristic of the outdoor scene. In an attempt to address this issue, previous methods learn a model to improve the realism of synthetic images. Differently from previous methods, this paper takes the first step towards purifying the naturalistic image to weaken the influence of light and convert the distribution of outdoor naturalistic image through a style transfer task to that of indoor synthetic image. This paper proposes therefore, a controlled neural style transfer network to preserve image structure, accelerate model convergence rate and adapt to multi-scale images. A mixed research approach (qualitative and quantitative) was adopted for the experiments carried out to demonstrate the possibility of purifying naturalistic images of complex distribution. Qualitatively, it compares the proposed method with baseline methods across several indoor and outdoor scenes of the LPW dataset. While quantitatively it evaluates the purified images by training models for gaze estimation on cross-dataset. Results show a significant improvement over using raw naturalistic images and when compared with the baseline methods.