Blind motion image deblurring using canny edge detector with generative adversarial networks / Idriss Moussa Idriss
Blind motion image deblurring has been investigated widely in recent years. Many methods and scheme shave been proposed so far, with respectto edge-preserving. Edge information transfers the essential details of an image which is a primary factor impacting the visual effect. Edge-preserving is an im...
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
2021
|
| Subjects: | |
| Online Access: | http://studentsrepo.um.edu.my/14436/ http://studentsrepo.um.edu.my/14436/1/Idriss_Moussa.pdf http://studentsrepo.um.edu.my/14436/2/Idriss_Moussa.pdf |
| Summary: | Blind motion image deblurring has been investigated widely in recent years. Many methods and scheme shave been proposed so far, with respectto edge-preserving. Edge information transfers the essential details of an image which is a primary factor impacting the visual effect. Edge-preserving is an important attribute during the process of image restoration. The main objective of deblurring is to generate a good approximation of theory original image from the blurry image. However, blind motion deblurring has remained challenging task for image processing and computer vision. Most of the existing algorithms rely on MAP (Maximuma Priori) and VB (Variational Bayesian)which are based on deterministic and stochastic methodologies, respectively, to estimate the blur function. MAP and VB both rely on particular assumptions to find the sources of the blur which make it difficult to use the edge-preserving treatment during the deblurring process. Therefore, including an edge-preserving treatment in the deblurring process would overcome these barriers as edges are the essential attribute of image information. This study proposes a combination approach of Canny edge detector with generative adversarial networks (GANs) to reconstruct the blurry image with edge-preserving without prior knowledge of the blurred image. The proposed method takes the blurred image with its detected edge, enhances it using Canny edged etectorasan input, and producesa corresponding detected restored sharp edge with its image, evaluated by the GANs. Then there stored sharp image is compared against the ground truth (sharp) image. Experiment s are conducted using the GoPro dataset. The proposed combined method has achieved good deblurring with edge-preserving results based on the evaluation metrics used.
|
|---|