【作者】 任静；

【导师】 苗启广；

【作者基本信息】 西安电子科技大学 ， 计算机应用技术， 2015， 硕士

【摘要】 2006年,Hinton提出了一种深层次结构的神经网络,再一次掀起了人工神经网络的热潮。深度学习是一个较为复杂的机器学习算法,目前在图像识别、语音识别等方面已经获得了很大的成功。深度学习使得机器能够模拟人类的视觉、听觉以及思考等活动,解决了许多比较复杂的模式识别方面的难题,使得人工智能的相关技术获得了非常大的进步。在当前的研究工作中,经常将深度学习与各类的实际应用结合起来,这是极其重要并且具有现实意义的。本文在研究图像去噪问题的基础上,给出了一个基于卷积神经网络的深度网络。该网络区别于传统的神经网络,由四个子网络构成,每个子网络有不同的隐层数目。每个网络中输入图像经过数次卷积操作,对输入图像进行不断滤波得到特征映射图,对这些特征映射图进行全连接得到输出图像。为了能够更好的说明本文给出的深度网络是如何对图像进行去噪的,将图像的去噪问题转换成一个学习问题,通过学习输出图像与标签图像的误差来修正网络中每层的权重,使得该深度网络能够对图像进行去噪。另外对该深度网络中输入的设置、网络层数的选择、训练迭代次数的设置、训练样本的大小、卷积层的构建以及构建模型的时候参数的初始化和学习率的设置问题等对网络的影响进行了分析。对于本文给出的深度网络的训练,当前一个子网络训练好之后,将其参数作为后一个网络的初始化,再对后一个子网络进行训练。卷积操作能够使得图像变小,而图像去噪问题在保证去除图像噪声的同时,图像的大小也不能变。在深度网络训练的过程中,每次迭代对输入图像中随机选一个点为中心,进行卷积,得到的图像来覆盖输出图像相应部位。本文设计了三个实验来验证所给出的去噪算法的性能。实验一设计了一幅512?512大小的图像,对该图像的不同部位加入高斯噪声,构成训练样本集。实验二和实验三分别对lena图像和woman图像的局部和整体加入高斯噪声。从主观和客观两个方面分析去噪结果,并与之前的方法作对比,实验结果较好。更多还原

【Abstract】 In 2006, Hinton proposed a neural network of deep structure. It sets off a boom in artificial neural networks once again. Deep learning is a relatively sophisticated machine learning algorithm. It has been a great success in the image recognition and speech recognition. Deep learning allows the machine to simulate human visual and thinking, and other activities. It solves many complex pattern recognition problems, which makes the relevant technology of artificial intelligence to get a lot. The combination of deep learning and a variety of practical applications is a very important research work.On the basis of studying the problem of image denoising, this thesis presents a deep network based on convolutional neural network. The network is different from the traditional neural network. It consists of four sub networks, each of which has a different number of hidden layers. The input image is filtered to obtain the feature maps through the convolution operation. Then these feature maps are all connected to the output image.In order to explain the deep network is how to denoise. The image denoising problem is converted into a learning problem in this thesis. Then the deep network learns the error of the output image and label image to modify the weight of each layer in the network, so that it can be used for image denoising. In addition, this thesis introduces the setup of the network, the number of layer, the setting of the training iteration number, the size of the training sample, the construction of the convolution layer, and the setting of the parameters of the model.In the training of the deep network, after the current one sub network training is good, the parameter is initialized as a next network. Then a next sub network is trained. Convolution operation can make the image become small. Image noise must be removed, while the size of the image cannot be changed. In each iteration of training, this thesis selects a region of the input image, which randomly selects a point as the center. It gets the image cover the corresponding parts of the output image through the convolution.In this thesis, three experiments are designed to verify the performance of the proposed algorithm. Experiment one designs an image of 512?512. The different blocks of the image are added to Gauss noise. Experiment two and experiment three respectively add Gauss noise to the local and the overall of the lena image and woman image. The results are analyzed from two aspects of subjective and objective, the results are better than the previous methods.更多还原