基于深度学习的图像去噪算法-计算机科学与技术专业论文.docx

Submitted in total ful?lment of the requirements for the degree of Master in Computer ScienceDeep Learning Based Image Denoising AlgorithmYangwei WuSupervisor Prof. Liqing ZhangS????? ?? E????????? I?????????? ??? E????????? E?????????? S??????? J??? T??? U?????????S???????, P.R.C????Jan. 21, 2015基于深度学习的图像去噪算法摘要图像去噪是图像处理的基本问题之一，经过多年的研究，目前的图像去噪 算法对于噪声强度较弱时已经能够达到很好的性能。然而在高噪声的环境下， 由于信噪比低，图片内可利用的信息不足，这些算法的性能会迅速下降。基于神经网络的图像去噪算法通过学习自然图像的统计特性，借助外部信 息来为去噪过程服务，十分适合高噪声环境下的去噪任务。然而近年来深度学 习的研究表明，传统的基于 S 形函数的神经网络存在着一定的局限性，这些局 限性限制了算法性能的提高。本文以神经网络的工具为基础，结合图像去噪问题的特性详细探讨了网络 参数的设置以及训练的技巧。并提出使用线性整流函数代替 S 形函数作为神经 网络隐藏层的激活函数以进一步增强神经网络解决去噪问题的能力。实验表明 通过更好地识别自然图像中的模式，我们的模型和传统的使用 S 形单元的模型 相比可以在更短的时间内达到更好的效果。在与现有的其它优秀去噪算法的比 较中，我们的算法取得了十分有竞争力的结果，尤其在高噪声环境下性能十分 优异，在消除噪声的同时很好地保留了图像中的细节信息。关键词： 图像去噪神经网络深度学习线性整流函数i —Deep Learning Based Image Denoising AlgorithmABSTRACTImage denoising is one of the basic problems in image processing. Last decade, current image denoising algorithms have been able to achieve quite good performance when the noise is weak. However, when the noise level is high, due to the low signal to noise ratio and the lack of useful information in the given image, the performance of these algorithms decreases rapidly.Neural networks are very suitable for the image denoising problem in high noise level by learning statistical properties of natural images. However, recent studies on deep learning have shown that neural networks based on the traditional sigmoid func- tion have its own limitations which limit its performance.In this paper, based on the tool of neural networks, we investigate the setting of the networks and the training procedure according to properties of the image denois- ing problem. We then propose to adopt the recti?er linear (ReL) function instead of the Sigmoid function as the activation function of hidden layers to further enhance the ability of neural network on solving image denoising problem. Our experiments show that by better capturing patterns in natural im