Research On Super Resolution Reconstruction From Surveillance Images

There is high demand for more details from the high-resolution image in the fields of remote sensing imaging, medical diagnosis and public security, etc. However, the low resolution of the image acquisition device limits the definition of digital images. Image super-resolution reconstruction is an effective way to solve this problem. It can reconstruct one or more clear high-resolution images from low-resolution images with subpixel non-redundant information.The process of image super-resolution reconstruction can be divided into three basic parts:image registration, interpolation reconstruction and image restoration. Based on the analysis of super-resolution reconstruction technique, this thesis starts from the three basic parts and centers on non-uniform interpolation algorithm which bases on norm convolution, then do some research on object tracking, image registration and modulation transfer function (MTF) estimation, etc. The main contributions are as follows:1) MeanShift object tracking algorithm based on color and edge information is proposed. In the video surveillance images, motion pattern between object and background is not consistent generally. Improved MeanShift object tracking algorithm is proposed for the problem in the thesis. It captures object from each frame and makes motion relationship between images from local motion to global motion. Compared with classic MeanShift object tracking algorithm, the improved algorithm which characterizes the object on color and edge information improves the accuracy of the tracking results and reduces the iteration times of registration algorithm.2) Improved six parameters Keren registration algorithm is proposed. Classic Keren algorithm which based on rigid motion model can not present the motion between the images under some condition, the proposed algorithm change the motion model to affine model with six parameters. In the case that the image rotates largely and zooms in or out, the proposed algorithm can still get more accurate results.3) Knife-edge based blind image restoration method is proposed. The thesis reduces the image noise by non-local mean filter on the stage of image restoration, and estimates the MTF by knife-edge method to make the blur identification, then restores the reconstructed image. Compared with other image restoration methods, the proposed method can enhance the definition of image obviously.Numerical and visual experiments have shown that the proposed algorithm can make images much clearer and obtain better performance over some classic algorithms.