| Summary: | 碩士 === 國立成功大學 === 電機工程學系 === 86 === Abstract Texture segmentation is an old and difficult problem
as yet. TextureSegmentation is similar to texture classification
except that the paternsof texture segmentation are pixels.
Unlike texture classification, the patternsare samples of
textures. Therefore, the difficulty of texture segmentaiton is
rather higher than texture classification. In this thesis, we
introduce a new feature, extrema fensity, extraced fromwavelet
coefficients (filtered images). According to our experiments, we
cansee that extrema density is more robust than the feature of
Low's energymeasure wich is widely used in multichannel texture
analysis. In fact, it isan encouraging result. Recently genetic
algorithms (GAs) have advanced theoretically and have been
successfully applied in many practical aspects. one of their
applicationareas is the optimization problem. On this thesis, we
applied GAs to search the subset, which results in the best
segmentaiton results, form original feature space. For the
objective function of GAs' part, we propose a measure,spatial
separabvility measure, to estimate the actual error rate. The
performance of segmentation obtained using the proposed measure
approximatesto that gained using the actual error rate. In our
experiments, the separable wavelet filters with tree-structured
and pyramid-structured transform and nonseparable wavelet
filters with tree-structured and pyramid-structured transform
are compared. The segmentation results without feature
selection, with feature seleciton using real error as the
fitness and with feature selction using spatial separability
measure as the fitness value are also compared. Experiments show
that the nonseparablewavelet filters with tree-structure
outperform other structures. Feature selection not only reduces
the computation load but also improves the segmentaion results.
|
|---|
