Learning Recursive Filters for Low-Level Vision via a Hybrid Neural Network
In this paper, we consider numerous low-level vision problems (e.g., edge-preserving filtering and denoising) as recursive image filtering via a hybrid neural network. The network contains several spatially variant recurrent neural networks (RNN) as equiv
- PDF / 4,627,450 Bytes
- 17 Pages / 439.37 x 666.142 pts Page_size
- 37 Downloads / 186 Views
2
UC Merced, Merced, USA {sliu32,mhyang}@ucmerced.edu Dalian University of Technology, Dalian, China [email protected]
Abstract. In this paper, we consider numerous low-level vision problems (e.g., edge-preserving filtering and denoising) as recursive image filtering via a hybrid neural network. The network contains several spatially variant recurrent neural networks (RNN) as equivalents of a group of distinct recursive filters for each pixel, and a deep convolutional neural network (CNN) that learns the weights of RNNs. The deep CNN can learn regulations of recurrent propagation for various tasks and effectively guides recurrent propagation over an entire image. The proposed model does not need a large number of convolutional channels nor big kernels to learn features for low-level vision filters. It is significantly smaller and faster in comparison with a deep CNN based image filter. Experimental results show that many low-level vision tasks can be effectively learned and carried out in real-time by the proposed algorithm.
1
Introduction
Recursive filters, also called Infinite Impulse Response (IIR) filters, are efficient algorithms that account for signals with infinite duration. As such, recursive implementations are commonly exploited to accelerate image filtering methods, such as spatially invariant/variant Gaussian filters [1–3], bilateral filters [4] and domain transforms [5]. However, few methods are developed based on recursive formulations for low-level vision tasks mainly due to the difficulty in filter design. Recently, several deep CNN based methods have been proposed for low-level vision tasks [6–10]. A convolutional filter can be considered as equivalent to a finite impulse response (FIR) filter. Unlike IIR filters, it is easier to design FIR filters at the expense of using more parameters to support non-local dependency. In deep CNNs, Xu et al. [7] approximate a number of edge-preserving filters by a data-driven approach which uses hundreds of convolutional channels to support spatially variant filtering or large (up to 16 × 16) kernels to support global convolution. In spite of using a large number of parameters, this model does not present local image structures well. Furthermore, it is difficult to extend the deep Electronic supplementary material The online version of this chapter (doi:10. 1007/978-3-319-46493-0 34) contains supplementary material, which is available to authorized users. c Springer International Publishing AG 2016 B. Leibe et al. (Eds.): ECCV 2016, Part IV, LNCS 9908, pp. 560–576, 2016. DOI: 10.1007/978-3-319-46493-0 34
Learning Recursive Filters for Low-Level Vision via a Hybrid NN
(a) smoothing
(b) denoising
(c) inpainting
561
(d) color interpolation
Fig. 1. Several applications of the proposed algorithm. (a) Approximation of relative total variation (RTV) [11] for edge-preserving smoothing. (b) Denoising. (c) Restoration of an image with random 50 % pixels occluded. (d) Restoration of an image with only 3 % color informations retained. Several applications of the proposed algorithm. (Colo
Data Loading...
