Color Image Reconstruction with Perceptual
Compressive Sensing
Jiang Du, Xuemei Xie
∗
, Chenye Wang, and Guangming Shi
School of Artificial Intelligence
Xidian University, Xi’an 710071, China
∗
Email: xmxie@mail.xidian.edu.cn
Abstract—We propose a novel compressive sensing framework
for color images. Recently, compressive sensing (CS) has gain its
popularity with the development of deep learning. To our best
knowledge, existing methods all deal with RGB images channel by
channel. This brings redundancy of measurements. In this paper,
we do a breakthrough work. Instead of recovering RGB images
channel by channel uniformly, we adopt non-uniform sampling
in different channels in YCbCr color space. The luminance
component takes up more measurements while the other channels
take up less in the proposed framework. It greatly enhances
the performance on CS for color images. Moreover, perceptual
loss gives a powerful ability to better capture the structure
information. We give the measurement rate at 2% as an example
in the experiments, and the results show the proposed method
outperforms all the existing methods with better structure of
images.
Index Terms—compressive sensing; color image reconstruc-
tion; perceptual loss; deep learning
I. INTRODUCTION
Compressive sensing (CS) [1] [2] [3] [4] [5] image recovery
is a classical problem in the field of computer vision. Tradi-
tional methods [6] [7] [8] [9] [10] [11] to solve CS problem are
mostly based on physical-driven approaches, such as greedy
algorithms [6] [7], convex optimization algorithms [8] [9], and
iterative algorithms [10] [11]. Since one set of measurements
can be obtained from different scene images, this is an ill-
posed problem and requires a lot of time consumption.
In recent years, deep learning technology plays an important
part in the field of computer vision. Deep neural networks
accelerate CS recovery and bring CS to a new height. Mousavi,
Patel, and Baraniuk [12] employ stacked denoising autoen-
coders (SDA) to recover signals from undersampled measure-
ments. This is the first time that deep learning technology is
employed on CS tasks. ReconNet [13] and DeepInverse [14]
use convolutional neural networks (CNN) to reconstruct the
signal from random Gaussian measurements. DeepCodec [15]
and Adp-Rec [16] break the random Gaussian measurements
and use learned measurements. All these above are block-
based methods, which cause block effect. FCMN [17] is firstly
introduced to remove block effect, and achieves state-of-the-
art results. Perceptual CS [18] makes a total difference that
greatly enhances the structure of the recovered images. To our
best knowledge, all the existing CNN-based CS methods are
based on single channel. When it comes to color images, they
have to work channel by channel uniformly. This brings a lot
of redundancy.
(a) Existing state-of-the-art result
(b) Proposed
Fig. 1. The proposed method gives a powerful enhancement on image
reconstruction compared to existing state-of-the-art method FCMN [17], with
both extremely low measurement rate at 2%.
In this paper, we propose a novel CS framework, which
deals directly with color images. This powerfully enhances
the performance of CS on color images. Inspired by human
vision system, we use the YCbCr color space [26] instead of
RGB. The non-uniform measurement network gives the image
more concentration on the intensity channel. This makes our
method better recover the structure of images. As is shown
in Fig. 1, the results of proposed method greatly outperform
the existing state-of-the-art method with pleasant visual effect
under the same measurement rate.
The major contribution of this paper is that, YCbCr color
model and non-uniform sampling are employed to directly deal
with color images. Besides, perceptual loss [19] is used in each
channel to better preserve the structure of images.
The outline of this paper is as follows. Related work is
introduced in Section 2. In Section 3, we present the main
technical design of the proposed framework. We show the
experimental results with analysis in Section 4. Section 5
draws the conclusion of this work.
2018 24th International Conference on Pattern Recognition (ICPR)
Beijing, China, August 20-24, 2018
978-1-5386-3787-6/18/$31.00 ©2018 IEEE 1512