免训练的卷积神经网络图像处理研究

近年来,卷积神经网络被大量的使用于图像处理领域,其特点在于能通过大量的图像样本进行训练,从而有效地获取图像的先验概率分布模型。研究卷积神经网络的结构在无需训练的情况下也可以获取图像的先验概率模型这一特性,通过设计特定结构的卷积神经网络,实现了较好图像的去噪和超分辨率处理;设计了一种基于梯度方向和梯度距离的图像质量评估方法,此方法能定量的评价生成图像的好坏,可用于衡量卷积神经网络的图像处理效果。

Recently,convolutional neural network has been widely applied in image processing area,the feature of it is that obtaining the prior probability distribution after network trained with a bundle of images.This paper researchs on the structure of convolutional neural network which could get the prior probability distribution of image without network training,designs a special convolutional neural network structure,achieving excellent denoise and super resolution processing in image,and provides a method base on gradient orientation and gradient magnitude to assess blind image quality,which could quantitative estimate image quality,then prove the performance of image processing with convolutional neural network.