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 with large receptive fields applied directly to the original color image. When training a deep model for these purposes however, it is quite difficult to generate edge-preserving images without undesirable color differences or boundary artifacts. To overcome these obstacles, we supervise intermediate outputs from different functional components during training. For example, 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 image decompositions without any special design, specialized auxiliary training data, 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.