Classification with wavelet transforms

Beltran et al. (50) succeeded in classifying 172 Chilean wines according to the type of grapes (cabernet sauvignon, merlot, and carmenere). First, phenolic compound chromatograms were developed with FIPLC-DAD. Second, features were extracted from the chromatographic data with different feature extraction techniques, like discrete Fourier transform and Wavelet transform. Finally, next to other different classification techniques, LDA and QDA were applied. From CV, both methods were found to result in acceptable correct classification rates without statistically significant difference between both rates. 

There are few possible strategies of library compression. Each of them has its own advantages and drawbacks. The most efficient method of data set compression, i.e. Principal Component Analysis (PCA), leads to use of global features. As demonstrated in [15] global features such as PCs (or Fourier coefficients) are not best suited for a calibration or classification purposes. Often, quite small, well-localized differences between objects determine the very possibility of their proper classification. For this reason wavelet transforms seem to be promising tools for compression of data sets which are meant to be further processed. However, even if we limit ourselves only to wavelet transforms, still the problem of an approach optimally selected for a particular purpose remains. There is no single method, which fulfills all requirements associated with a spectral library s compression at once. Here we present comparison of different methods in a systematic way. The approaches A1-A4 above were applied to library compression using 21 filters (9 filters from the Daubechies family, 5 Coiflets and 7 Symmlets, denoted, respectively as filters Nos. 2-10, 11-15 and 16-22). 

ABSTRACT In the discrete wavelet transform approach, a choice of wavelet has a direct impact on the decomposed image, which indicates that the selection of the wavelet is closely related to the detection performance. Since a choice of standard wavelets, e.g. Daubechies wavelets, Coiflets, biorthogonal wavelets etc., may not guarantee efficient discrimination of fabric defects, some researchers suggested methods based on a construction of wavelets adapted to the detection or classification of these defects. We propose a novel method to design adaptive wavelet filters. These filters are constructed to minimize /2-norm of the undecimated discrete wavelet transform of the defect free textile with the aim to enhance the wavelet response in the defect region. Examples show efficiency in the fault detection. 

Therefore, other option from various methods of classification with extracted features such as wavelet transform... 

T. Chang and C. Kuo, Texture analysis and classification with treestructured wavelet transform, IEEE Trans. Image Processing, vol. 2, pp. 429-441, Oct. 1993. 

---