Improved clustering approach for junction detection of multiple edges with modified freeman chain code
Image processing framework of two-dimensional line drawing involves three phases that are detecting junction and corner that exist in the drawing, representing the lines, and extracting features to be used in recognizing the line drawing based on the representation scheme used. As an alternative to...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2015
|
| Subjects: | |
| Online Access: | http://eprints.utm.my/id/eprint/77663/1/HaswadiHasanPFC2015.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Image processing framework of two-dimensional line drawing involves three phases that are detecting junction and corner that exist in the drawing, representing the lines, and extracting features to be used in recognizing the line drawing based on the representation scheme used. As an alternative to the existing frameworks, this thesis proposed a framework that consists of improvement in the clustering approach for junction detection of multiple edges, modified Freeman chain code scheme and provide new features and its extraction, and recognition algorithm. This thesis concerns with problem in clustering line drawing for junction detection of multiple edges in the first phase. Major problems in cluster analysis such as time taken and particularly number of accurate clusters contained in the line drawing when performing junction detection are crucial to be addressed. Two clustering approaches are used to compare with the result obtained from the proposed algorithm: self-organising map (SOM) and affinity propagation (AP). These approaches are chosen based on their similarity as unsupervised learning class and do not require initial cluster count to execute. In the second phase, a new chain code scheme is proposed to be used in representing the direction of lines and it consists of series of directional codes and corner labels found in the drawing. In the third phase, namely feature extraction algorithm, three features proposed are length of lines, angle of corners, and number of branches at each corner. These features are then used in the proposed recognition algorithm to match the line drawing, involving only mean and variance in the calculation. Comparison with SOM and AP clustering approaches resulting in up to 31% reduction for cluster count and 57 times faster. The results on corner detection algorithm shows that it is capable to detect junction and corner of the given thinned binary image by producing a new thinned binary image containing markers at their locations. |
|---|