基于BP神经网络的数字图像修复

考虑到BP神经网络非凡的学习能力和非线性映射能力,提出了利用BP神经网络修复数字图像.由于一般的BP神经网络收敛速度较慢,且易陷入局部极小,产生振荡现象.因此考虑在梯度下降算法的基础上引进动量因子,结果发现收敛速度加快、振荡现象减轻.该方法根据待修复区域的边界寻找相似块,利用相似块周围像素数据得到BP神经网络的权值和阈值.试验表明:文中的方法相对于利用偏微分方程(如BSCB方法)速度要快,而且具有更大的ISNR.

Image inpainting based on BP neural network

BP neural network used to repair digital images has be proposed, considering the traits of BP neural network which has an extraordinary capacity for learning and the non-linear mapping. Because ordinary BP neural network converge at a slower rate, fall in to a local minimum easily, and produce oscillatory behavior, introduction of momentum on the basis of the gradient descent algorithm is considered. It get faster convergence speed, smaller oscillation phenomena. Similar blocks will be found in accordance with the border of areas to be restored. With similar block pixel data, BP neural network get its weights and thresholds. The experiment results show this model spend less time than the model utilized partial differential equation （PDE） （for example BSCB）, and this model has larger ISNR than PDE model.