Video Motion Deblurring Base On Target Tracking

When obtaining image and video, it is inevitable to be degradation, which will make the image information be with noise and blurring in real life. The blurred image is always accompanied by information losing and inaccurate, which perplex the issue. Losing the high frequency information increase the difficulty to analyze and recognize. This thesis discuss image deblur mainly for video blur caused by objects in motion capered by fixed camera. How to deblur the moving object and how to deal with all different objects together is the focus of this thesis.This thesis extracts the blurring object firstly. Using brilliant image as the priori and the constraint of point spread function to establish the regularization function to gain the point spread function. Let the super Laplace function fit the gradient histogram of the brilliant image. Transform the progress of image deblurring to the question of minimize the energy function according to the Bayes theory. The result has better evaluation index both in subjective and objective. What is more, it is close to the fast motion deblur algorithm in the time. The specific process of the deblurring consist of four parts, extracting the blurred foreground, estimating the point spread function's initial value, estimating the accurate value and deblurring the image.1) This thesis describes two different tracking algorithms, Compressive Target and Multi-Instance Learning. The experiment result shows that the Multi-Instance Learning is better to extract the foreground. Then transform the spatially changing point spread function to the spatially invariant point spread function.2) Using the speed-up robust features to match the moving object between two sequential frames to compute the motion vector. The initial value of point spread function is estimated by the scaling relation of the exposure time and the time between two sequential frames. This method get the better point spread function and be more robust to the noise compared to the frequency domain method.3) Combining the constraint of point spread function and brilliant image gradient prior to establish the regularization function. Add the character prior to make the character clear.4) Super Laplace regularization gets less ringing effect compared to the classics Richardson-Lucy algorithm and objective indicator natural image quality evaluator is littler.