Abstract
Image processing and pixel-wise dense prediction have been advanced byharnessing the capabilities of deep learning. One central issue of deeplearning is the limited capacity to handle joint upsampling. We present a deeplearning building block for joint upsampling, namely guided filtering layer.This layer aims at efficiently generating the high-resolution output given thecorresponding low-resolution one and a high-resolution guidance map. Theproposed layer is composed of a guided filter, which is reformulated as a fullydifferentiable block. To this end, we show that a guided filter can beexpressed as a group of spatial varying linear transformation matrices. Thislayer could be integrated with the convolutional neural networks (CNNs) andjointly optimized through end-to-end training. To further take advantage ofend-to-end training, we plug in a trainable transformation function thatgenerates task-specific guidance maps. By integrating the CNNs and the proposedlayer, we form deep guided filtering networks. The proposed networks areevaluated on five advanced image processing tasks. Experiments on MIT-AdobeFiveK Dataset demonstrate that the proposed approach runs 10-100 times fasterand achieves the state-of-the-art performance. We also show that the proposedguided filtering layer helps to improve the performance of multiple pixel-wisedense prediction tasks. The code is available athttps://github.com/wuhuikai/DeepGuidedFilter.