| Summary: | 碩士 === 國立臺灣大學 === 電信工程學研究所 === 94 === The automatic recognition of images has been a researched topic over decades yet still a difficult task to accomplish. Similar as the process of human perception, the first step of automatic recognition should be image segmentation to segment different homogeneous patches into regions. However, like the automatic recognition, segmentation remains a yet satisfactorily solved problem. This is due to the uncertainty and complexity nature of images. By the aid of various image processing techniques and pattern recognition analysis, this problem may reduced to some less complicated level, thus helps to achieve preferable result.
We first introduce two excellently performed algorithms of color image segmentation up to date, and made some modification on the structural phase and improved their efficiency while preserving resolution and accuracy. Moreover, we experimented on adding some special approaches as pre-processing, tested on specific images which are difficult to segment and get very good results.
From another aspect to look upon this problem, since the recognition and also the segmentation so difficult a problem to solve, why not ask for a little assistance of human specification?
In the latter part of the thesis, interactive object extraction is briefly covered and a new system we proposed is tested. With simple user specification, we could solve the extraction problem by a simpler and faster version of solution. This reduction of complexity should thanks to the prior knowledge given by human specification.
Both the automatic image segmentation and interactive object extraction take the advantage of multi-scale structure.
|
|---|
