Effect of masking techniques on computational complexity reduction of scale invariant feature transform
The Scale Invariant Feature Transform (SIFT) is algorithm use in feature detection and description, it is famous and has dominated the research community. The SIFT is a standard compared to others. Notwithstanding its sturdiness, SIFT has limitation of computational complexity, this has pose great l...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2015
|
| Online Access: | http://psasir.upm.edu.my/id/eprint/65602/1/FK%202015%20150IR.pdf http://psasir.upm.edu.my/id/eprint/65602/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The Scale Invariant Feature Transform (SIFT) is algorithm use in feature detection and description, it is famous and has dominated the research community. The SIFT is a standard compared to others. Notwithstanding its sturdiness, SIFT has limitation of computational complexity, this has pose great limitation to real time and on-line implementations. Effort has been done by researchers to improve on SIFT performance because of it robustness, the computational cost was reduced to some extent but the distinctiveness cannot be compared with others(Descriptors). Thus, the aim of this research is to propose masking techniques to images, by eliminating areas with no or sparse keypoints in the feature extraction process, thereby reducing the computational cost of SIFT as the dominant descriptor in computer and robotic vision. Performance of the proposed approach was able to reduce the computational time to 47.27% at 0.7 Threshold with 17.94% in keypoint reduction. However, the masked SIFT was used in place categorization for performance evaluation. The evaluation was conducted using two classifiers in pattern recognition on the classification of Royal Institute of Technology-Image Data for Robot Localization (KTH-IDOL2) database. The classifiers are Nearest Neighbor (NN) and Multilayer Perceptron (MLP). Comparison of the categorization and classification accuracy produced 70% with the nearest neighbor(NN) and improved to 80% with Multilayer perceptron. In conclusion, the proposed approach was able to improved the computational time of SIFT higher than the recent published work in literature. |
|---|