基于卷积神经网络的图像模糊去除

发布时间：2018-04-23 11:12

本文选题：图像模糊去除 + 卷积神经网络　；参考：《安徽大学》2017年硕士论文

【摘要】：摄像机已经渗透到人们生活的方方面面,图像作为摄像机的产物,为人们传递信息提供了另一种重要途径。然而由于种种原因,例如拍摄物体的运动、对焦不准确、光照条件的不足等因素,造成拍摄得到的图像是不清晰的。这些质量退化的图像往往不能满足人类的需求,严重时会造成一定的经济损失。去除模糊、复原图像、提高图像质量成为人们研究工作关注的重点。本文主要针对去除图像模糊与噪声、恢复清晰图像方面开展研究。首先,介绍了基于卷积神经网络的图像模糊去除研究的相关理论工作。分析图像退化模型的基础上,讨论了几种常用的求解模型,并详细介绍了卷积神经网络的基本原理及面向图像模糊去除的卷积神经网络模型。其次,依据前文的卷积神经网络的理论知识,分析了现存的深度学习模糊去除方法的不足,详细的介绍了本文设计的高频信号保持且可快速模糊去除的快速卷积神经网络模型(fast CNN,FCNN)。在该网络模型训练的过程中,对高频图像进行傅里叶域上的梯度预处理,通过实施傅里叶域模糊去除的预处理得到一个初始的清晰图像。接着将该初始图像小块作为输入,相应的真实清晰图像小块作为标签训练FCNN,得到从模糊图像到潜在清晰图像的映射函数,实现基于该模型的模糊去除。预处理过程中,添加了梯度约束的高斯模型与提高平滑度的约束项,将图像的先验特征直接用于模糊去除,得到一个较鲁棒的初值,为后续训练FCNN奠定基础。FCNN模型由四层卷积层及激活函数构成,这样设计的目的在于移除图像模糊的同时,降低时间复杂度。实验结果表明,基于FCNN的图像模糊去除方法相对于其他深度学习模糊去除方法,在有效的移除图像模糊的基础上,能够更好的保持图像的高频纹理信息,同时降低了时间复杂度。最后,针对FCNN模型严重依赖梯度约束预处理导致FCNN方法自适应较低的缺陷,提出增加模型深度方式得到改进的FCNN模型—β-FCNN,并结合最小二乘方滤波预处理的方法实现一个新的去模糊方法,提高了自适应性。首先考虑到卷积核大小为1×1的卷积层能够在增加模型深度的情况下,最小化网络训练参数的优点,对FCNN模型改进,得到β-FCNN模型;其次,考虑到最小二乘方滤波算法能够使滤波后的图像边缘得到增强。基于该特性,本文将最小二乘方滤波算法与β-FCNN模型结合(称之为β-FCNN方法)用于图像模糊去除。网络训练过程中,将最小二乘方滤波预处理后的图像块作为输入,对应的清晰图像块作为标签,训练该模型,得到从模糊图像到潜在清晰图像的映射函数,实现基于该模型的图像模糊去除。实验结果表明,β-FCNN方法相对于其他深度学习方法,有效的移除图像模糊的基础上,在一定范围内,有较强的自适应性。
[Abstract]:Cameras have penetrated into all aspects of people's lives. As a product of cameras, images provide another important way for people to transmit information. However, due to a variety of reasons, such as the motion of the shooting object, focusing inaccurate, insufficient lighting conditions and other factors, resulting in the image is not clear. These degraded images often can not meet the needs of human beings, and will cause certain economic losses. Removing blur, restoring images and improving image quality have become the focus of research. This paper focuses on removing image blur and noise and restoring clear image. Firstly, the theoretical work of image blur removal based on convolution neural network is introduced. Based on the analysis of image degradation model, several common solving models are discussed, and the basic principle of convolution neural network and the convolution neural network model for image fuzzy removal are introduced in detail. Secondly, according to the theoretical knowledge of convolution neural network, the shortcomings of the existing deep learning fuzzy removal methods are analyzed. The fast convolution neural network model designed in this paper is a fast fast convolution neural network model which can keep high frequency signal and can be removed quickly. In the process of training the network model, the gradient preprocessing of the high-frequency image is carried out in the Fourier domain, and an initial clear image is obtained by the pre-processing of the fuzzy removal in the Fourier domain. Then, the original image block is used as input and the corresponding real clear image block is used as label to train FCNN, and the mapping function from blurred image to latent clear image is obtained, and the fuzzy removal based on the model is realized. In the process of pretreatment, the Gao Si model with gradient constraint and the constraint item to improve the smoothness are added, and the prior features of the image are directly used to remove the blur, and a more robust initial value is obtained. The model is composed of four layers of convolution layer and activation function. The purpose of this design is to remove image blur and reduce the time complexity. The experimental results show that compared with other depth-learning fuzzy removal methods, the image fuzzy removal method based on FCNN can effectively remove the image blur, and can better keep the high-frequency texture information of the image. At the same time, the time complexity is reduced. Finally, in view of the disadvantage of FCNN model which depends heavily on gradient constraint preprocessing, the FCNN method is less adaptive. An improved FCNN model- 尾 -FCNN with increasing the depth of the model is proposed, and a new deblurring method is implemented by combining the least square filter preprocessing method, which improves the self-adaptability. Firstly, considering that the convolution layer with a convolution kernel size of 1 脳 1 can minimize the advantages of network training parameters while increasing the depth of the model, the 尾 -FCNN model is obtained by improving the FCNN model. Considering that the least square filter algorithm can enhance the edge of the filtered image. Based on this characteristic, the least square filtering algorithm and 尾 -FCNN model are combined in this paper for image blur removal. In the process of network training, the preprocessed image block of least square filter is taken as input, and the corresponding clear image block is used as label to train the model, and the mapping function from blurred image to latent clear image is obtained. Image blur removal based on this model is realized. The experimental results show that compared with other depth learning methods, the 尾 -FCNN method can effectively remove image blur and has strong adaptability in a certain range.
【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】