Single Scene Image Editing Based on Deep Intrinsic Decomposition

Hao Sha, Yue Liu, Kai Lu, Chenguang Lu, Hengrun Chen, Yongtian Wang

Abstract: Intrinsic decomposition is an inherent problem in computer graphics and computer vision, which decomposes an image into a reflectance image and a shading image. Through the processing of reflectance and shading, it can achieve the image scene editing effect consistent with the vision of the real scene, which endows application potentials to augmented reality. In this paper, we propose a convolutional neural network structure that includes shared and independent encoder to the independent decoder as well as several different loss functions for training based on the assumption of intrinsic decomposition and differences in datasets. To address the shortcomings of the existing synthetic dataset, we reconstruct the new synthetic data and train our network on the synthetic and real datasets in sequence. We quantitatively and qualitatively evaluate our intrinsic decomposition results on the IIW dataset, and the result shows that they outperform those of existing methods. We also perform image editing based on our deep intrinsic decomposition on images of real different scenes and obtain satisfactory visual results.