基于梯度匹配增强的单帧监控视频超分辨率方法

A Single Surveillance Image Super-Resolution Method Based on Gradient Matching

由于采集环境和设备的限制,监控视频中如人脸等重要目标区域面积较小,插值放大技术会使得目标模糊,影响后续处理效果。本文研究单帧监控视频的超分辨率技术,处理流程分为训练和增强两个阶段。在训练阶段,通过对相关图像库的训练得到由低分辨率图像块和高分辨率图像块对构成的训练集;在增强阶段,通过输入的低分辨率图像构造搜索矢量在训练集中匹配搜索找到高分辨率信息。为了提高搜索准确率,将梯度匹配增强到搜索矢量构造过程中,从而能在训练集中找到更符合真实结构分布的高分辨率信息,提高超分辨率增强的效果。实验结果表明,该方法得到的高分辨率图像边缘更锐利,视觉效果更好。

In surveillance video, the area of important objects such as faces often appears to be small, for the limits of capture conditions and devices. Interpolating images usually results in a blurring of edges and image details, then it is impossible for the tasks such as the recognition of faces. This paper presents the technique of super-resolution in surveillance scenarios, the process of which includes training and enhancing. In the training phase, we get the training data which are made up of low-frequency patches and high-frequency patches. In the enhancing phase, we construct the search vector from the low-resolution image patch and get the high resolution information from the training data through matching. In order to improve the veracity of searching, we introduce the gradient matching to construct the search vector, so we can get the most similar patches and improve the enhancing effect. The experiment shows that this method preserves fine details, such as edges, generates believable textures and gives good results.