Image Super-Resolution based on Multi-Convolution Neural Network
Guodong Jing
Distance Education Center
China Meteorological Administration Training Center
Beijing China
jinggd@cma.cn
Yun Ge
Department of computer teaching and research
University of Chinese Academy of Social Sciences
Beijing China
gyunsus@163.com
Abstract—In recent years, convolutional neural network
method has been widely and successfully applied in the field of
image super-resolution. With the development of CNN
structure, the reconstruction algorithm based on CNN has also
been developed. However, in these reconstruction models, the
scale of convolution operator is single. This will greatly limit
the model's learning ability about input image and impact the
reconstruction effect. In order to improve the accuracy of the
convolution network to the input image, a reconstruction
method based on multi-scale convolution operator is proposed.
In this method, multi-scale convolution operators are set up in
each layer network to calculate multi-scale features of the
input image. Experiments show that this method can
effectively improve the accuracy of the reconstructed image in
detail.
Keywords- Super-Resolution, Deep learning, Neural network,
Deep Residual Networks
I. I
NTRODUCTION
Single image Super-resolution (SISR) is to restore a low
resolution image (LR) to a high resolution image (HR). This
problem has always been an active research topic in the field
of computer vision, and it is also a classic ill-posed problem
of mapping from low-dimensional data to high-level data. At
present, super-resolution reconstruction methods based on
single image can be divided into three categories:
interpolation based method, prior knowledge based method
and depth learning based method. The methods based on
interpolation include nearest neighbor, bilinear and bicubic
interpolation. These methods are simple and efficient, but the
quality of image reconstruction is limited. The method based
on prior knowledge is based on the cognition of the
researcher to the image features. The feature calculation way
is designed based on the image prior information and the
calculation method of reconstruction parameters is
determined by the optimized customized model. Finally the
image reconstruction is completed. Using this method can
get better reconstruction results, but it also takes a long time
to optimize the model. Although the convolution neural
network (CNN) method and the method based on prior
knowledge can improve the efficiency to a certain extent,
these methods will still be affected by the calculation
efficiency. The method based on deep learning can
automatically learn the mapping relationship between low
resolution image and high resolution image to complete the
reconstruction work. In these methods, the prior knowledge
used to guide image reconstruction does not need to be
customized manually, but acquired through learning.
In recent years, the methods based on deep learning have
been widely used in various fields. This kind of method has
also achieved very good results in image super-resolution
reconstruction. Dong et al. [1] is the first one to use the
convolutional neural network to complete the reconstruction
work (SRCNN), and get very good reconstruction results.
But this method also has some shortcomings. First of all,
SRCNN has too few network layers. It only uses three
convolution layers to complete the tasks of feature extraction,
feature mapping and image reconstruction. They claim that
too many convolution layers can not significantly improve
the reconstruction quality while reducing the computational
efficiency. However, the network with small depth usually
has a relatively small perceptive field, which makes it
difficult for the network structure to learn the large-scale
features of the image. The size of the network receptive field
will directly affect the content of context information, which
is often used to infer the details of that feature. A wide range
of receptive fields can provide more information for the
model. Compared with the deep network, the shallow
network contains very few nonlinear activation layers. And
the nonlinear mapping of network is just an important
mechanism for deep learning network to improve network
diversity. Later, Dong et al [2] proposed FSRCNN to
improve the SRCNN method. The network structure has a
deeper network level and faster convergence speed. But the
network with too deep layer will encounter the problem of
gradient disappearance and over fitting. In order to solve this
problem, Kim et al. Proposed a super-resolution
reconstruction method (VDSR) based on residual
convolution neural network [3]. The network model uses a
gradient truncation and jump connection mechanism to
alleviate the problem of over fitting and gradient
disappearance. In addition to increasing the number of
network layers, VDSR also speeds up the network training
process by improving the learning rate. They further
proposed using DRCN [4] to control the scale of model
parameters. Tai et al. [5] further proposed a deep loop
residual network (DRNN) by introducing residual learning
mechanism and cyclic structure, which uses the mechanism
of parameter sharing to reduce the number of parameters.