| Summary: | 碩士 === 中華大學 === 機械與航太工程研究所 === 90 === The object of this study is to use texture to class color images of clothes and packing papers. To achieve the objective, the following processing procedures are proposed. First, the color image is transformed from RGB model to other suitable model, and then one of the components is chosen as the gray-level image. Secondly, the gray-level image is decomposed into four child images by using wavelet transform technique. Thirdly, The two child images capable of detecting variations along columns (vertical edges) and variations along rows (horizontal edges) are used to generate 0-degree and 90-degree co-occurrence matrices, respectively. Fourth, The two co-occurrence matrixes are used to compute some of the standard features representing texture characteristics. The calculated features include angular second moment, entropy, contrast, homogeneity, dissimilarity, correlation, probability of a run length, and cluster tendency. After that, some of the distinguishable features are used as the inputs to the back-propagation neural network. Finally, according to the outputs of the BPN network, the test image can be classified.
To verify the usefulness of the proposed classification method, two groups of samples are used. The first group consists of nine packing papers with either different colors or textures. The second group consists of eight curtain cloths with either different colors or textures. The experimental results show that if both color and texture information are used in the classification process, the recognition rates are 97.855% and 97.842% for packing papers and cloths, respectively. If the mentioned two groups of samples are mixed, i.e., 17 classes to distinguish, the recognition rate is 97.184%. The classification results are satisfactory.
|
|---|
