Revisiting Deep Image Smoothing and Intrinsic Image Decomposition

We propose an image smoothing approximation and intrinsic image decomposition method based on a modified convolutional neural network architecture applied directly to the original color image. Our network has a very large receptive field equipped with at least 20 convolutional layers and 8 residual units. When training such a deep model however, it is quite difficult to generate edge-preserving images without undesirable color differences. To overcome this obstacle, we apply both image gradient supervision and a channel-wise rescaling layer that computes a minimum mean-squared error color correction. Additionally, to enhance piece-wise constant effects for image smoothing, we append a domain transform filter with a predicted refined edge map. The resulting deep model, which can be trained end-to-end, directly learns edge-preserving smooth images and intrinsic decompositions without any special design or input scaling/size requirements. Moreover, our method shows much better numerical and visual results on both tasks and runs in comparable test time to existing deep methods.