Intrinsic image extraction based on deconvolutional neural networks

Intrinsic image decomposition provides an important way to analyse the real characteristics of the objects in images. However, it is an ill-posed problem, where two outputs, shading and reflectance, are extracted from only one given image. Therefore extra constraints such as the consistence of color, texture and shape are always applied to solve this problem. Due to the superb feature extraction ability, deep learning neural networks improve the performance of many computer vision tasks, including intrinsic image analysis. However, the pixel-bypixel reconstruction of reflectance and shading is still a challenge for the traditional classification-oriented deep neural network. In this paper, we propose an end-to-end double stream neural network to reconstruct the reflectance and shading simultaneously. For the proposed neural network, features from different layers are applied for the reconstruction which can enhance the introduction of details. Meanwhile the constraint from double stream also can improve the accuracy. In the experiments, the proposed network shows its effectiveness based on the training and evaluation on a dataset of real images.