Textural feature extraction and optimization in wavelet sub-bands for discrimination of green tea brands

This study aimed to discriminate green tea brands with textural feature from wavelet sub-bands based on multi-spectral image. Firstly, 250 multi-spectral images of five brands tea were obtained from a common-aperture multi-spectral charged coupled device camera with three channels (550, 660 and 800 nm). Secondly, each image was converted into seven wavelet sub-bands images by wavelet pyramidal decomposition at second level. Then statistic textural features such as contrast, homogeneity, energy, correlation and entropy were calculated from grey level co-occurrence matrix (GLCM) of wavelet sub-bands image. 105 textural features were obtained by feature extraction way combined by wavelet transform and GLCM. Thirdly, statistical feature selection was used to optimize the number of textural feature. 11 characteristic features were selected from 105 original features through STEPDISC of SAS with high statistic significance. Discriminant functions were generated based on these 11 characteristic features. Perfect classification performance (100%) was obtained for samples both in training and prediction sets. It can be concluded that green tea brands can be effectively discriminated by texture analysis based on multi-spectral image.