Self-Supervised Dual Generative Networks for Edge-Preserving Image Smoothing

Huiqing Qi, Shengli Tan, Xiaoliu Luo

Abstract: Image smoothing aims to remove insignificant textures or perturbations while maintaining meaningful structures, which is a fundamental technique widely used in the vision and graphics fields. However, the performance of deep methods is often limited by ground-truth images and edge preservation ability in this task. To address the two issues, we proposed self-supervised dual generative networks (S2DGNet) with the Cauchy regularized variational model. The dual networks integrate information on textures and structures under the alternating direction method of multipliers (ADMM) framework, improving the model’s edge-keeping. In addition, we created a new smoothing dataset named the ECS dataset. Experiment results demonstrate the noticeable performance improvement of the S2DGNet over the existing solutions.