Paper Link:: http://jiaya.me/all_final_papers/motion_deblur_cvpr07.pdf
- Abhinaba Bala (2020701001), abhinaba.bala@research.iiit.ac.in
- Neel Ashok Mishra (2020701009), neel.mishra@research.iit.ac.in
- Ritam Basu (2020701005), ritam.basu@research.iiit.ac.in
- Jigyasu Khandelwal (2020702013), jigyasu.khandelwal@research.iiit.ac.in
- Image blur is caused due to object motion or camera vibration when the shutter is pressed. The relationship between the blur filter and the object boundary transparency has to be understood.
- An algorithm is to be developed to estimate the blur filter from the transparency values given a blurred image. This formulation is to be used to estimate a filter for general transparency present in an image.
- This will alow us to deblur an image irrespective of the cause of the blur, which is a first of its kind method (when) formulated by the authors.
- Motion blur is usually formulated as a linear image degradation process, given by I = L * f + n, where I, L and n represent the degraded captured image, unblurred (or latent) image, and the additive noise respectively. * is the convolution operator and f is the unkown linear shift-invariant point spread function.
- Conventional blind deconvolution methods estimate the blur filter from image intensities or gradients and deconvolve the blurred image.
- A shortcoming of this method is that it can't completely solve the probelm because the background may not undergo the same motion as the object, as the filter is defined only on the moving object.
- The relationship between image boundary transparency and the deblur filter is investigated and it is shown that a filter can be estimated from only the transparency values.
- An algorithm is developed to estimate the filter using a Maximum A Prosteriori (MAP) formulation with a suitable prior and likelihood on transparency.
- A formulation of general transparency is investigated and it is shown that the previous formulation is robust for any kind of motion blur.
- The algorithm is also very efficient since only image patches are used as input.
- Iterative optimization method for our MAP approach to recover the motion blur filter using transparency.
Employ conjugate gradient optimization and then Belief Propagation to estimate the filter - Deconvolution using Lucy-Richardson method.
- Use user-drawn strokes to collect the foreground and background samples for object motion blur.
Presentation Link:: https://github.com/barvin04/Image-Deblurring/blob/main/documents/Presentation_ProximaCentauri.pdf
Project Proposal Submission: 18th October
25th October: Analysis of Paper
31st October: Set up the starter code and dependencies
Mid Evaluation: 31st October
7th November: Implement MAP approach to recover the motion blur filter using transparency and solve 2-D object motion blur.
14th November: Implement generalized transparency for when the entire image is degraded.
19th November: Complete the codebase and prepare presentation.
Final Evaluation: 19th-25th November
No