A Study of Single Image Deblurring Using Gradient-Based Learning

• Main Authors: Kuan-Jen Chen, 陳冠任
• Other Authors: Jiunn-Lin Wu
• Format: Others
• Language: zh-TW
• Published: 2016
• Online Access: http://ndltd.ncl.edu.tw/handle/80866019246267214935

碩士 === 國立中興大學 === 資訊科學與工程學系 === 104 === Motion blur is one of the common defects in digital photography. While there is not enough light, we need to use long shutter speed for the good image quality. Take a photo with long shutter speed and without camera tripod, it may result in a motion blur image. To solve this problem, image deblurring is an active topic in computational photography and image processing fields. A blurred image can be modeled as unblurred image convolutes the movement of camera (or called the point spread function), if the kernel function of motion blur is assumed shift-invariant. Mostly, we do not have the point spread function. So we focus on the study of blind image deconvolution. We estimate point spread function accurately by using edge patch to enhance blurred image, but the deblurred image is still having some ringing artifact. In the study, we propose an efficient and adaptive image deconvolution method, it combines the edge patch enhancement and bilateral filter for the restoration of the burred images. We use edge patch to enhance image and bilateral filter to reduce ringing artifacts which is caused by noise and narrow edge. Finally, we apply image pyramid strategy to accelerate the process of image deblurring. We expect that the proposed method can not only achieve the result without ringing artifact, but also have a good image deconvolution result.