- Author
- Nazanin Sadat Hashemi and three co-authors, Islamic Azad University
- Published
- 23 October 2016
- Kind
- Review paper, arXiv preprint
- Original
- arxiv.org/abs/1610.07231
- Subject
- AI put to use
What it is
Four authors from Islamic Azad University, North Tehran Branch wrote a literature review. It covers the maths of matching measures, efficiency and invariance. It then covers applications such as faces, circuit boards and medical images.
What it says
- The paper defines template matching as finding where a template appears in an image.
- It says normalized cross-correlation ignores brightness changes but is sensitive to rotation and scale.
- It tabulates operation counts for measures such as normalized cross-correlation and sum of absolute differences.
- For scale and rotation invariance, it describes Fourier-Mellin and log-polar style transforms.
- It reports Jalil's Malaysian plate system: 92.78 percent character accuracy on 100 plates.
- It reports Sarraf's brain network work, where NCC and Dice did best.
- The paper concludes that matching should be paired with a decision-making step.
What evidence it gives
The paper summarizes methods and equations from earlier work. It includes a table of operation counts. It reports figures from cited studies, such as 92.78 percent character accuracy on licence plates.
Interests and limits
The paper is from 2016 and predates later deep learning advances in the field. It reports other groups' results and not its own tests. The brain imaging section leans on work by one cited group. Several sections give little detail on how results were measured.