摘要
传统通过预测点扩展函数的图像盲复原方法具有依赖于边缘的缺点,这限制其在平滑图像上的效果。提出一种基于卷积神经网络的图像复原方法,通过训练两个结构相同的网络Pre-Net和Iter-Net,分别实现模糊图像特征分解和重构清晰图像的目的,从而达到省略传统图像复原方法中预测点扩展函数等中间步骤,直接进行图像-图像的复原效果。实验证明,该方法具有良好的图像盲复原效果,在一定程度上克服传统图像依赖边缘特性的缺点。
Proposes a new method named Match-Map for image deblur which trains a network named Pre-Net and keeping another network named Iter-Net still whose structure is the same with the Pre-Net. Respectively, decomposes the blurred image into feature maps and then reconstruet it to clean image depending on the feature maps. The proposed deblur method can skip the traditional step of estimate the point spread function （PSF） of blurred images and deblur the images directly. Experiments show that this method has a good blind image restoration effect.
出处
《现代计算机》
2017年第6期111-114,119,共5页
Modern Computer
