基于卷积神经网络的图像超分辨率算法研究及GPU实现
发布时间：2018-07-21 20:02
【摘要】：随着显示技术的飞速发展,人们对高分辨率的图像、视频资源的需求也日益增长。然而由于图像采集设备的硬件限制,高分辨率的资源比较稀缺,并且高质量的图像在存储与传输上也有比较高的要求,而解决这种矛盾的一个非常好的方法就是图像超分辨率。目前在图像超分辨率领域,目前超分辨率重建效果较好的算法是基于学习的算法。这些算法都有着优秀的超分辨率效果,然而这些算法大都将注意力集中于超分辨率的重建效果上,在算法的效率上则不尽如人意。最近由于深度学习热潮的兴起,卷积神经网络被应用于图像超分辨率领域,这种使用卷积神经网络进行超分辨率的方法称为超分辨率卷积神经网络,也是一种基于学习的方法。超分辨率卷积神经网络由于其简单的前馈网络结构,算法无需求解复杂的优化问题,因此算法在效率上相比其它算法了有很大的提升。虽然超分辨率卷积神经网络在效率上已经超越了其它基于学习的算法,然而对于实时性要求较高的应用来说,超分辨率卷积神经网络的时效性还是不足以满足实时超分辨率的要求。针对超分辨率卷积神经网络的时效性问题,本文首先对小规模的超分辨率卷积神经网络进行优化,使得小规模的超分辨率卷积神经网络也能够取得良好的超分辨率效果。其次借助CUDNN库,将超分辨率卷积神经网络在GPU上实现,用以提高卷积神经网络在PC上的计算效率。最后,本文将GPU上的卷积神经网络进行封装,并应用在视频压缩仿真及增强处理系统中,使得系统能够实现实时的视频超分辨率功能。
[Abstract]:With the rapid development of display technology, the demand for high resolution images and video resources is also increasing. However, due to the hardware limitation of image acquisition equipment, high resolution resources are scarce, and high quality images are also required in storage and transmission. A very good method to solve this contradiction is image super-resolution. At present, in the field of image super-resolution, the algorithm of super-resolution reconstruction is based on learning. These algorithms have excellent super-resolution effects, but most of them focus on the super-resolution reconstruction effect, but the efficiency of these algorithms is not satisfactory. Recently, due to the rise of deep learning craze, convolutional neural network is applied to image super-resolution field. This method using convolution neural network for super-resolution is called super-resolution convolution neural network, and it is also a learning-based method. Due to the simple feedforward network structure of super-resolution convolution neural network, the algorithm does not need to solve complex optimization problems, so the efficiency of the algorithm is greatly improved compared with other algorithms. Although super-resolution convolution neural network has outperformed other learning-based algorithms in efficiency, however, for applications with high real-time requirements, The timeliness of super-resolution convolution neural network is not enough to meet the requirement of real-time super-resolution. Aiming at the time-efficiency of super-resolution convolution neural network, this paper first optimizes the small-scale super-resolution convolution neural network, so that the small-scale super-resolution convolution neural network can also obtain good super-resolution effect. Secondly, the super-resolution convolution neural network is realized on GPU with the help of CUDNN library, so as to improve the computing efficiency of convolution neural network on PC. Finally, this paper encapsulates the convolutional neural network on GPU and applies it to video compression simulation and enhancement processing system, which enables the system to realize real-time video super-resolution function.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TP183
本文编号：2136704
[Abstract]:With the rapid development of display technology, the demand for high resolution images and video resources is also increasing. However, due to the hardware limitation of image acquisition equipment, high resolution resources are scarce, and high quality images are also required in storage and transmission. A very good method to solve this contradiction is image super-resolution. At present, in the field of image super-resolution, the algorithm of super-resolution reconstruction is based on learning. These algorithms have excellent super-resolution effects, but most of them focus on the super-resolution reconstruction effect, but the efficiency of these algorithms is not satisfactory. Recently, due to the rise of deep learning craze, convolutional neural network is applied to image super-resolution field. This method using convolution neural network for super-resolution is called super-resolution convolution neural network, and it is also a learning-based method. Due to the simple feedforward network structure of super-resolution convolution neural network, the algorithm does not need to solve complex optimization problems, so the efficiency of the algorithm is greatly improved compared with other algorithms. Although super-resolution convolution neural network has outperformed other learning-based algorithms in efficiency, however, for applications with high real-time requirements, The timeliness of super-resolution convolution neural network is not enough to meet the requirement of real-time super-resolution. Aiming at the time-efficiency of super-resolution convolution neural network, this paper first optimizes the small-scale super-resolution convolution neural network, so that the small-scale super-resolution convolution neural network can also obtain good super-resolution effect. Secondly, the super-resolution convolution neural network is realized on GPU with the help of CUDNN library, so as to improve the computing efficiency of convolution neural network on PC. Finally, this paper encapsulates the convolutional neural network on GPU and applies it to video compression simulation and enhancement processing system, which enables the system to realize real-time video super-resolution function.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TP183
【参考文献】
相关期刊论文 前2条
1 浦剑;张军平;黄华;;超分辨率算法研究综述[J];山东大学学报(工学版);2009年01期
2 苏秉华,金伟其,牛丽红,刘广荣;超分辨率图像复原及其进展[J];光学技术;2001年01期
，本文编号：2136704
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2136704.html
