采用卷积神经网络的小幅文本图像重聚焦算法

针对文本图像拍摄过程中的散焦模糊问题，提出一种基于卷积神经网络的图像重聚焦算法．首先分析了传统的维纳滤波方法，并对其进行变形；然后将频域相除转化为循环卷积，并将该卷积核进行奇异值分解，从而将二维卷积转化为一维卷积．在重点考虑循环卷积、一维卷积核的基础上，设计出了一种新的卷积神经网络结构．该网络结构不但能够模拟维纳滤波的去散焦模糊过程，还能在不显式计算散焦模糊核的情况下恢复图像，并具有良好的抗噪声性能．同时，该卷积神经网络还具有收敛快、参数不敏感的良好特点．

Text image refocusing by using the convolutional neural network

We propose a new image refocusing algorithm based on the convolutional neural network. We analyze the traditional wiener filtering method, and derive it. We transform the frequency domain division into a cyclic convolution and decompose the kernel by SVD. After that, we design a new structure of the convolutional neural network by cyclic convolution and one-dimensional convolution. This network can not only simulate the defocusing process of wiener filtering, but also restore the image without a kernel, and have good anti-noise performance. At the same time, the convolutional neural network is fast-convergent and parameter insensitive.