A novel fuzzy co-occurrence matrix for texture feature extraction

• Main Authors: Yutthana Munklang, Sansanee Auephanwiriyakul, Nipon Theera-Umpon
• Format: Book Series
• Published: 2018
• Subjects: Computer Science; Mathematics
• Online Access: https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=84880735297&origin=inward; http://cmuir.cmu.ac.th/jspui/handle/6653943832/52439
• Institution: Chiang Mai University

Texture analysis is one of the important steps in many computer vision applications. One of the important parts in texture analysis is texture classification. This classification is not an easy problem since texture can be non-uniform due to many reasons, e.g., rotation, scale, and etc. To help in this process, a good feature extraction method is needed. In this paper, we incorporate the fuzzy C-means (FCM) into the gray level co-occurrence matrix (GLCM). In particular, we utilize the result from FCM to compute eight fuzzy co-occurrence matrices for each direction. There are four features, i.e., contrast, correlation, energy, and homogeneity, computed from each fuzzy co-occurrence matrix. We then test our features with the multiclass support vector machine (one-versus-all strategy) on the UIUC, UMD, Kylberg, and the Brodatz data sets. We also compare the classification result using the same set of feature extracted from the GLCM. The experimental results show that the features extracted from our fuzzy co-occurrence matrix yields a better classification performance than that from the regular GLCM. The best results on validation set using the features computed from our fuzzy co-occurrence are 77%, 95%, 99.11%, and 98.44% on the UIUC, UMD, Kylberg, and Brodatz, respectively, whereas those on the same data sets using the features from the gray level co-occurrence are 53%, 85%, 82.81%, and 95.31%, respectively. The best result on the blind test set of Brodatz data set using our fuzzy co-occurrence is 92.19%, whereas that from the gray level co-occurrence is 85.74%. Since the blind test data set is a rotated version of the training data set, we may conclude from the experiment that our features provide better property of rotation invariance. © 2013 Springer-Verlag Berlin Heidelberg.