Fast End-to-End Trainable Guided Filter

Abstract

Dense pixel-wise image prediction has been advanced by harnessing the capabilities of Fully Convolutional Networks (FCNs). One central issue of FCNs is the limited capacity to handle joint upsampling. To address the problem, we present a novel building block for FCNs, namely guided filtering layer, which is designed for efficiently generating a highresolution output given the corresponding low-resolution one and a high-resolution guidance map. Such a layer contains learnable parameters, which can be integrated with FCNs and jointly optimized through end-to-end training. To further take advantage of end-to-end training, we plug in a trainable transformation function for generating the task-specific guidance map. Based on the proposed layer, we present a general framework for pixel-wise image prediction, named deep guided filtering network (DGF). The proposed network is evaluated on five image processing tasks. Experiments on MIT-Adobe FiveK Dataset demonstrate that DGF runs 10-100 times faster and achieves the state-of-theart performance. We also show that DGF helps to improve the performance of multiple computer vision tasks.

Multitask bilateral learning for real‐time image enhancement

“…We implemented the proposed algorithm using Tensorflow . We compared deep bilateral learning (DBL), deep guided filtering (DGF), the proposed single bilateral learning (SBL), which trains a single image processing operator, and MBL, which trains multiple image processing operators simultaneously in global and local image processing operators.…”

“…We resized the full HD images to 512 × 512 in the training phase for efficiency, and full HD images were used in the testing phase. The proposed network (SBL) and the conventional networks (DBL, DGF) were compared with the expert A, B, and C datasets, respectively. To train various image processing operators from the same input image simultaneously, expert A, B, and C datasets were set to the ground truth, and multitask learning was conducted by the MBL network for 240 epochs.…”

“…Recently, an approach combining joint upsampling and the DNN has been proposed for real‐time image processing . Deep guided filtering applies the guided filter to the deep learning structure.…”

“…Recently, an approach combining joint upsampling and the DNN has been proposed for real‐time image processing . Deep guided filtering applies the guided filter to the deep learning structure. It downsamples the input image from full resolution to low resolution and estimates the low‐resolution output image through a context aggregation network .…”

“…However, because DNNs usually require a large number of parameters and high computational cost for high performance, these methods are not suitable for real‐time applications. In recent years, algorithms combined with both joint upsampling and deep learning approaches have been actively researched . These algorithms process images by DNN in low resolution and restore them to full resolution through a joint upsampling operator.…”