基于卷积神经网络的图像复原方法研究

发布时间：2018-07-10 04:34

本文选题：图像去模糊 + 卷积神经网络　；参考：《江南大学》2017年硕士论文

【摘要】：随着虚拟现实技术和移动互联网的不断发展,图像在人类获取和传递信息的过程中扮演着重要角色。在日常生活中,图像和视频不断的充斥在我们的视野中。由于人眼对图像的清晰度比较敏感,清晰度低的图像会给观看者带来不舒服的感受。所以相关显示设备技术才得以不断发展来满足人们的需求。然而现实生活中存在一些不可避免的因素会导致获得的图像失真,例如图像采集过程中摄像机未聚焦、相机抖动等。所以说利用失真图像来还原出原始图像是一件非常重要的事情,具有很大的实际意义。实现这个过程的技术我们称为图像复原技术,是目前比较热门的研究方向。本论文利用卷积神经网络模型在图像复原上的优势,研究了图像复原涉及的两个热门问题图像去模糊和图像超分辨率重建。本文研究内容体现在下面几个方面:(1)提出一种基于卷积神经网络的图像去模糊模型。避免了传统的图像去模糊算法对模糊图像先验知识的依赖。在点扩散函数未知的情况下,通过网络训练学习输入的模糊图像与目标清晰图像二者之间的非线性映射关系,实现图像去模糊。通过实验,在选择网络参数时,对该图像去模糊方法在性能和时间上做出权衡,再将最佳的参数应用于该模型上。实验表明该模型优于传统图像去模糊算法。(2)提出一种基于混合神经网络的图像去模糊模型。混合神经网络由卷积神经网络与BP神经网络(Back propagation Neural Network)组成,二者分步实现图像复原。首先,通过训练卷积神经网络提取退化图像有效感知特征,再将提取的特征向量作为BP神经网络的输入来训练BP神经网络,从而实现图像去模糊。实验表明该方法在小尺度的模糊核上的复原效果显然优于现有方法,但是当模糊核的尺度超过23×23的情况下,复原效果明显下降。(3)提出一种改进的基于卷积神经网络的图像超分辨率重建模型。基于卷积神经网络的图像超分辨率重建模型包含3个卷积层,它们的作用分别为提取图像块特征、非线性映射和重建。本文通过增加网络的层数,改变卷积层中滤波器的数量,改变卷积层中滤波器尺寸等来改进基于卷积神经网络的超分辨率重建技术。改进的卷积神经网络包含四个卷积层和一个下采样层。下采样层采用最大、中值、最小池三池联合的方法,不仅可以有效提取图像的质量感知特征而且有利于提高训练效率。实验结果表明,该模型可以有效逼近真实的高分辨率图像。
[Abstract]:With the development of virtual reality technology and mobile Internet, images play an important role in the process of obtaining and transmitting information. In our daily life, images and videos are constantly filled with our vision. Because the human eye is sensitive to the sharpness of the image, the low-definition image will bring uncomfortable feelings to the viewer. Therefore, the related display equipment technology can be continuously developed to meet the needs of people. However, there are some unavoidable factors in real life, such as camera unfocusing, camera jitter and so on. So it is very important to restore the original image by using the distorted image, which has great practical significance. The technology to realize this process, which is called image restoration technology, is a hot research direction at present. In this paper, the advantage of convolution neural network model in image restoration is used to study the image de-blurring and super-resolution reconstruction of two hot problems involved in image restoration. The main contents of this paper are as follows: (1) an image de-blurring model based on convolution neural network is proposed. It avoids the dependence of the traditional image de-blurring algorithm on the priori knowledge of the blurred image. When the point diffusion function is unknown, the nonlinear mapping relationship between the inputted fuzzy image and the target clear image is studied by network training, and the image deblurring is realized. Through experiments, when the network parameters are selected, the performance and time of the image de-blurring method are weighed, and the best parameters are applied to the model. Experiments show that the proposed model is superior to the traditional image de-blurring algorithm. (2) A hybrid neural network based image de-blurring model is proposed. Hybrid neural network is composed of convolution neural network and back propagation neural network. Firstly, the effective perceptual feature of degraded image is extracted by training convolution neural network, and then the extracted feature vector is used as the input of BP neural network to train BP neural network to realize image de-blurring. The experimental results show that the restoration effect of this method on the small scale fuzzy kernel is obviously better than that of the existing method, but when the scale of the fuzzy kernel is more than 23 脳 23, (3) an improved image super-resolution reconstruction model based on convolution neural network is proposed. The image super-resolution reconstruction model based on convolution neural network consists of three convolution layers which are used to extract image block features nonlinear mapping and reconstruction respectively. In this paper the super-resolution reconstruction technique based on convolution neural network is improved by increasing the number of layers in the network changing the number of filters in the convolution layer and changing the size of the filter in the convolution layer. The improved convolution neural network consists of four convolution layers and one downsampling layer. The method of maximum median and minimum pool is adopted in the lower sampling layer which not only can extract the image quality perception features effectively but also can improve the training efficiency. Experimental results show that the model can approach real high resolution images effectively.
【学位授予单位】：江南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】