CN113362226A

CN113362226A - Image super-resolution reconstruction based on deep learning

Info

Publication number: CN113362226A
Application number: CN202110668604.3A
Authority: CN
Inventors: 陆绮荣; 吴止境; 卢子任
Original assignee: Guilin University of Technology
Current assignee: Guilin University of Technology
Priority date: 2021-06-16
Filing date: 2021-06-16
Publication date: 2021-09-07

Abstract

The invention provides an image super-resolution reconstruction method based on deep learning. The method comprises the following specific steps: (1) and (5) super-resolution reconstruction of the image. Learning from a large amount of training data by using a Convolutional Neural Network (CNN) to obtain a mapping relation model in a low-definition picture and a high-definition picture, and performing super-resolution reconstruction processing on the images in the test set based on the CNN by using the mapping relation model. (2) And (3) after the step (1) is completed, obtaining a reconstructed image, using the reconstructed image as a low-definition image, defining a generator network and a discriminator network by adopting a Keras deep learning framework, performing feature extraction by using a VGG19 pre-training model to construct an SRGAN model, and finally performing SRGAN-based super-resolution reconstruction processing on the images in the test set. (3) And (3) after the step (2) is completed, the super-resolution reconstruction of the image based on the deep learning can be realized. The method has the advantages of convenient use, high response speed, high measurement precision and simple structure, and improves the field detection speed.

Description

Image super-resolution reconstruction based on deep learning

Technical Field

The invention relates to a method for reconstructing super-resolution images, which belongs to the field of digital image processing, the field of image super-resolution reconstruction application and wide application, and has important application in the aspects of military affairs, medicine, public safety, computer vision and the like.

Background

The concept of super-resolution was first proposed in the optical field, which refers to a process in which one tries to recover data outside the diffraction limit. But with the advent of the digital age, we are becoming more and more closely related to consumer electronics.

Unlike before, people touch sensing media such as pictures and videos through modern mature consumer electronics instead of observing optical objects in reality with naked eyes.

Thus, the quality of the digital media we are exposed to is affected by a variety of factors due to the limitations of the medium. If limitations such as transmission bandwidth, sampling device performance, and too low original resolution of rendering are received, only low-definition images lower than the display resolution of the viewing device can be obtained, and these low-definition images often cannot meet the requirements of viewing experience. As in the application of video surveillance in a public security system, the image quality degrades for a variety of reasons, but high-definition images are often the basis for police personnel to obtain critical information in a surveillance mission. As is common to most people, the most common one is faced with the situation where the resolution of the display device for viewing does not match the resolution of the media content. Nowadays, with rapid development in the field of consumer electronics, the display resolution of TVs, high-resolution displays, thin and light notebook, ipads, and even small and medium-sized mobile phone screens used by people has reached a very high level, and the development of these display devices towards high resolution is promoted to some extent by the appearance of 8k, 4k and other streaming media. But in practice media images are still provided, usually at a lower resolution, due to limitations in transmission bandwidth, server capacity, etc. Too little picture resolution is a rather poor viewing experience for high definition display devices. In a display device with a high unit pixel density, the physical size displayed by a low-resolution image will be too small to be viewed.

The image super-resolution technology is an effective technical means introduced to meet the above requirements. The image super-resolution technology tries to enlarge the size of a digital image by a certain multiple, fits the information of the enlarged missing pixels through different algorithms and mathematical means, and supplements the information of the newly added pixels, so that a super-resolution image with a display effect superior to that of a low-resolution original image to a certain extent is obtained to meet the requirements of people.

Disclosure of Invention

The invention aims to overcome the defects of the existing method and provides an image super-resolution reconstruction method based on deep learning.

The method comprises the following specific steps:

(1) super-resolution reconstruction of images essentially uses a priori relationship to fit unknown information. And learning from a large amount of training data by using a Convolutional Neural Network (CNN) to obtain a corresponding relation between the low-definition pictures and the high-definition pictures, namely a model of a mapping relation. And performing super-resolution reconstruction processing on the images in the test set based on the CNN by using the mapping relation model.

(2) And (3) after the step (1) is completed, obtaining an image after super-resolution reconstruction based on the convolutional neural network, using the image as a low-definition image, defining a generator network and a discriminator network by adopting a Keras deep learning framework, performing feature extraction by using a VGG19 pre-training model to construct an SRGAN model, and finally performing the super-resolution reconstruction processing based on the SRGAN on the test concentrated image.

(3) And (3) after the step (2) is completed, the super-resolution reconstruction of the image based on the deep learning can be realized.

The method is simple and convenient, and the used method is convenient to use, high in response speed, high in measurement precision and simple in structure, so that the problem of complicated extraction of the defect characteristics of the underground pipeline is fundamentally solved, and the field detection speed is greatly improved.

Drawings

Fig. 1 is an overall flow chart of the SRCNN.

Fig. 2 is a structure diagram of a three-layer network of the SRCNN.

FIG. 3 is a graph comparing the results of the classical algorithm and SRCNN.

FIG. 4 is a flow chart of SRGAN method model construction.

FIG. 5 is a graph of the results of combining CNN with GAN.

Detailed Description

The proposed framework for extracting the fracture characteristics of the underground sewage pipeline is shown in the attached drawings of the specification. The algorithm herein comprises three parts: (1) learning from a large amount of training data by using a Convolutional Neural Network (CNN) to obtain a corresponding relation between a low-definition picture and a high-definition picture, performing super-resolution reconstruction on the picture for one time (2), outputting a high-definition result obtained in the step (1) as low-resolution input (3) obtained in the step (2), defining a generator network and a discriminator network by adopting a Keras deep learning framework, performing feature extraction by using a VGG19 pre-training model to construct an SRGAN model, and performing super-resolution reconstruction processing on the picture in a test set based on the SRGAN.

Claims

1. The image super-resolution reconstruction based on the deep learning comprises the following specific steps:

(1) the super-resolution reconstruction of the image essentially utilizes a priori relationship to fit unknown information; learning from a large amount of training data by using a Convolutional Neural Network (CNN) to obtain a corresponding relation between a low-definition picture and a high-definition picture, namely a model of a mapping relation; performing super-resolution reconstruction processing on the images in the test set based on the CNN by using a mapping relation model;

(2) after the step (1) is completed, obtaining an image after super-resolution reconstruction based on a convolutional neural network, using the image as a low-definition image, defining a generator network and a discriminator network by adopting a Keras deep learning framework, performing feature extraction by using a VGG19 pre-training model to construct an SRGAN model, and finally performing super-resolution reconstruction processing based on the SRGAN on the test concentrated image;