利用BP神经网络实现监控图像盲复原

Blind Restoration of Monitoring Image Based on BP Neural Network

运用大量的真实图像与退化图像数据样本训练BP神经网络,然后再用训练好的网络进行实际的图像复原,方法可以实现真正的盲复原。但是,由于很难解决如何采集训练素材的问题,一直以来得不到实际的应用。针对监控图像应用的特点,通过设计实验,在原始拍摄场地对已有清晰图片进行拍摄,得到的退化图像经配准后和原始清晰图像共同组成训练图像对,解决了训练网络的材料的来源问题。实验表明,复原图像在视觉上和定量分析上都获得了良好的效果。

By using numbers of real images and degraded images for neural network training, and then using the trained network for image restoration, real blind rehabilitation can be realized. But it＇s difficult to resolve the proplem of acquiring training materials, so there has been no practical application. In this paper, according to the characteristics of monitoring image, a new experiment is designed. Images of prepared photos are taken in the original shooting venue and then used to train the neural network. Thus, the problem of lacking training materials is solved. Experiments show that this method has a satisfactory outcome both in visual impression and quantitative analysis.