基于快速卷积神经网络的图像去模糊

发布时间：2019-05-14 20:41

【摘要】：针对基于深度学习的图像去模糊方法无法有效地保留高频纹理信息,易产生振铃效应,且时间复杂度较高的问题,提出基于卷积神经网络(CNN)的图像去模糊方法.该方法设计了一种高频信号保持且可快速去模糊的快速CNN模型(FCNN).在此基础上,首先对高频图像进行傅里叶域上的预处理,通过实施傅里叶域去模糊的预处理得到一个初始的清晰图像;然后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,实现基于该训练网络的去模糊.定性和定量实验结果表明,文中方法利用CNN参数共享的特点,减少了网络训练过程中大量的学习参数;相对前人基于深度学习的去模糊方法,该方法对模糊图像在保持图像纹理细节恢复的同时使计算复杂度得到显著降低.
[Abstract]:In order to solve the problem that the image deblurring method based on depth learning can not effectively retain the high frequency texture information, it is easy to produce ringing effect, and the time complexity is high, an image deblurring method based on convolution neural network (CNN) is proposed. In this method, a fast CNN model (FCNN). With high frequency signal preservation and fast deblurring is designed. On this basis, the high frequency image is preprocessed in Fourier domain, and an initial clear image is obtained by deblurring the Fourier domain. Then the initial image block is used as the input, and the corresponding real clear image block is used as the label to train the FCNN, to get the mapping function from the blurred image to the potentially clear image, and the deblurring based on the training network is realized. The qualitative and quantitative experimental results show that the method in this paper makes use of the characteristics of CNN parameter sharing to reduce a large number of learning parameters in the process of network training. Compared with the previous deblurring methods based on depth learning, this method can significantly reduce the computational complexity of blurred images while maintaining the restoration of image texture details.
【作者单位】：安徽大学媒体计算研究所;安徽大学计算智能与信号处理教育部重点实验室;
【基金】：国家自然科学基金(61301295) 安徽省科技攻关计划(1604d0802004) 安徽省自然科学基金(1408085MF113,1608085QF129)
【分类号】：TP391.41

【相似文献】