-
Notifications
You must be signed in to change notification settings - Fork 0
GSoC 2014 Siva Prasad Varma Implementation of Image Registration Module
Name: Siva Prasad Varma Chiluvuri
Email ID: sivapvarma@gmail.com
Telephone: +91-9557020908
Time Zone: UTC+5:30 (IST)
IRC Handle: sivapvarma@irc.freenode.net
GitHub: sivapvarma
Skype Username: shivaprasadvarma
Twitter: sivapvarma
Home Page: http://sivaprasadvarma.in/
Blog: http://skimage-gsoc.blogspot.in/
GSoC Blog RSS Feed: http://skimage-gsoc.blogspot.com/feeds/posts/default
University: Indian Institute of Technology Roorkee
Major: Electronics & Communication Engineering with Specialization in Wireless Communication
Current Year: Fifth (Final)
Expected Graduation Date: June 2014
Degree: Integrated Dual Degree (Bachelors + Masters)
- (WIP)Shape Contexts - https://github.com/scikit-image/scikit-image/pull/921
- (WIP)use plt.subplots() in examples - https://github.com/scikit-image/scikit-image/pull/926
Image Registration is an important part in many image processing applications like Image Stitching, Remote Sensing and Super Resolution. This project aims to add a registration sub-module to scikit-image that implements both spatial domain and transform domain registration methods
Image Registration methods fall under two Categories: Spatial domain & Transform domain methods. Spatial domain methods operate in the image domain, matching intensity patterns or features in Images. Thus spatial methods can further be classified as dense registration and sparse registration. In dense registration methods, similarity metric like SSD, SAD and Normalized Cross Correlation(NCC) can be used with uni-modal images while Mutual Information used for multimodal images (for eg. MR and CT in medical imaging). In sparse registration(feature based methods), the Homography is estimated from correspondences. The correspondences can be found using a Brute Force Matcher (skimage.feature.match_descriptors) or a Nearest Neighbour solver. RANSAC is used to remove the outliers. Transform domain methods find the transformation parameters for registration of the images while working in the transform domain. Phase correlation and log-polar registration fall under transform domain methods.
- dense registraion using Mutual Information, NCC, SAD, SSD similarity metrics
- sparse registration
- log-polar registration
- Phase correlation
A majority of these methods are implemented in either [1] or [2].
- Discuss implementation details and decide the communication and reporting protocol with mentors
- Read papers in references thoroughly
- Decide on the API of
registrationsub module. - Understand
imreg&supreme.registercodebases
- Implement dense registration
- SSD, SAD, NCC
- Mutual Information
- Implement dense registraion
- Mutual Information
- Implement Sparse registration
- Least square estimation via direct method.
- Add documentation & examples
- Code clean up
- Documentation & adding examples
- Buffer time for backlogs
- Implement Sparse Registration.
- least square estimation via Iterative method.
- Implement Phase Correlation[5]
- Implementation of log-polar registration[6]
- Code optimization, cythonize if needed.
- Add examples and documentation
- code clean up
- Buffer period to complete backlog
- Image Stitching algorithms
I am free for the entire summer except for three days in june when I travel from my college to home and three days in july when I'll go to Bangalore. I'll try to schedule these mostly in weekends, if not possible I'll code during weekends.