Motion Deblurring For Natural Images

Low-level vision problem is one of the most basic and challenging problems.It has received much attention by researchers in many fields.As one branch of low-level vision problems,image deblurring aims to solve the following issue:given low resolution images,tries to give the according high-resolution images,which is recognizable for humans.Image deblurring has been discussed by researchers in computer vision,signal processing,pattern recognition for many years,but still cannot be well solved.In this dissertation,I tried to give some new views for image deblurring problems using learning free methods and learning based methods.1?Motion blur estimation using dark channel and convolution operators.In this section,a robust algorithm for motion blur estimation was proposed.Some facts were used as prior to guide the deblurring process.The results shown that this algorithm gives comparative results.2?Motion blur estimation using7)₀ regularization and convolution operators.In this section,a robust algorithm for motion blur estimation was proposed.Some observations were used this information as priori to guide the deblurring process.The results shown that this algorithm gives comparative results.3?Blind motion deblurring using data-driven method.Traditional motion blur kernel estimation algorithms often have expensive computational cost,and they cannot fully utilize the advantages of big data.In this section,we proposed a blind mot-ion blur estimation model using cycle generative adversarial networks tries to solve this problem.The results shown that this model gave comparative results.