Abstract:
Wavelet transform is a multi-scale technical analysis which is popular for
image analysis and it has been applied to many aspects of it. The method is to divide
the original image into multi-scales through the convolution with the mother wavelet,
where the translation and scaling can acquire subimages. Then, the texture feature will
be extracted from subimages, which can be used to describe the image in more detail.
Previous research which has discriminated texture image through wavelet
transform and which only used a small number of Brodatz textures. It had an accuracy
rate of more than 90%. However, in trials of the entire 112 texture categories, it was
found that the accuracy rate decreased to approximately 70%. Because of this, this
research proposed the wavelet transform and dynamic feature selection, which chooses
the feature for the adoption on the automatic discrimination of the texture category
through the vote from each feature. There are 5 features implemented in this research,
i.e. norm-1(average energy), standard deviation, average residual, entropy (log energy)
and maximum probability.
This research used two experimental methods on Brodatz textures. The first
experiment was tried on 20 texture categories. The trial with the tree structure wavelet
transform and the dynamic feature selection acquired an accuracy rate of 97.9%. The
second experiment was tried on 112 texture categories. The outcome of the trial on the
pyramid wavelet transform and tree structure wavelet transform had accuracy rates of
77.7% and 74% respectively. However, upon adopting the tree structure wavelet
transform and dynamic feature transform the accuracy rate increased to 87%. It is
apparent that the tree structure wavelet transform and dynamic feature selection can
improve the classification result.