Incoporating fuzzy clustering into gray level co-occurrence matrix

การผสมผสานการจัดกลุ่มแบบฟัซซีกับเมทริกซ์การเกิดร่วมของระดับสีเทา

Name: Yutthana Munklang

ThaSH: Fuzzy logic

Classification :.DDC: 629.89

ThaSH: Computer engineering

Abstract: Texture is one of the most important image characteristics. Texture analysis is also one of the important steps in many computer vision applications. Particularly, texture classification, a major part in texture analysis, is not an easy problem since texture can be non-uniform due to many reasons, such as rotation and scaling. A good feature extraction method is, therefore, needed in order to help in this process. In this research, we incorporated the Fuzzy C-Means (FCM) clustering into the Gray Level Co-occurrence Matrix (GLCM). In particular, we computed the Fuzzy Co-Occurrence Matrices (FCOM) from the result of FCM for each cluster. There are fourteen features, i.e. contrast, correlation, energy, homogeneity, variance, sum average, sum variance, entropy, sum entropy, difference variance, difference entropy, information measure of correlation, and maximal correlation coefficient, computed from each FCOM. We then tested our features with the multiclass support vector machine (one-versus-all strategy) on the Brodatz, Kylberg, UIUC, and the UMD gray levels texture data sets. We also compared the classification result using the same set of feature extracted from the GLCM. The best validation set using the features computed from our FCOM are 100%, 100%, 100%, and 100% on the Brodatz, Kylberg, UIUC, and UMD, respectively, whereas those on the same data sets using the features from the GLCM were 100%, 99.55%, 92%, and 99%, respectively. The experimental resultsshowed that the features extracted from our FCOM provided a better classification performance than those from the regular GLCM. Next, we developed a breast abnormalities detection system using the feature extracted from FCOM on the mini-MIAS public data set. The abnormalities considered were calcification (CALC), well-defined/circumscribed masses (CIRC), speculated masses (SPIC), and architectural distortion (ARCH). The feature set computed from the GLCM was also used for the purpose of comparison. The best blind test data set results of the ARCH, CALC, CIRC, and SPIC detection from the feature set computed from our FCOM were 100% with 9.46 false positives per image (FPI), 100% with 3.39 FPI, 81.25% with 18 FPI, and 90% with 13.72 FPI, respectively. The results of the same abnormalities detection from the feature set extracted from the GLCM were 100% with 9.46 FPI, 89.47% with 10.81 FPI, 68.75% with 6.78 FPI, and 70% with 4.45 FPI, respectively. From the experimental results, our system provided better performance than other existing methods in ARCH and CALC detection. Our results from our system were similar to the results from those methods used in SPIC and CIRC detection. However, our system did not require any preprocessing or region of interest (ROI) selection at all. We also implemented our FCOM texture features in the synthetic aperture radar (SAR) automatic target recognition (ATR) on MSTAR public release data set. The SAR image classification is one of the most challenging problems because of the difficult characteristics of SAR images. The MSTAR consists of three types of military vehicles, i.e., T72 tank, BMP2 armored personnel carriers (APCs), and BTR70 APCs. In classification task, we used MSVM and the radial basis function (RBF) network as classifiers. Moreover, we implemented the ensemble average as a fusion tool. The best recognition result was at 97.94% correct detection from the fusion of the 20 best FCOM with RBF network models (the 10 best RBF network models at d = 5 and the other 10 best RBF network models at d = 10) while the best fusion result of FCOM with MSVM was at 95.37% correct detection (the 10 best MSVM models at d = 5 and the other 10 best MSVM models at d = 10). We also compared the fusion detection result using the same set of features extracted from the GLCM. The experimental results from FCOM were better than those from GLCM in all cases. Finally, we also evaluated our features on the Outex, USPTex, Biomass, MondialMarmi, and NewBarkTex color texture data sets. In this case, the feature was combined from each color channel. We also compared the classification result using the same set of feature extracted from the GLCM and Fuzzy Color Levels Co-occurrence Matrix (FCLCOM). The FCLCOM was computed from the result of FCM where it clustered all color channels altogether. Furthermore, we also compared the classification result using the feature extracted from the Color Levels Co-occurrence Matrix (CLCM). The best validation set classification results using the features computed from our FCOM were 91.91%, 97.03%, 100%, 100%, and 97.55% on the color texture data sets, respectively. The results on the same data sets using the features extracted from the GLCM were 97.06%, 96.86%, 100%, 100%, and 97.55%, respectively. The same data sets using the features extracted from the FCLCOM were 89.71%, 89.53%, 100%, 100%, and 90.80%, respectively, whereas those on the same data sets using the features extracted from the CLCM were 96.32%, 96.51%, 100%, 100%, and 96.32%, respectively. Therefore, we concluded that the features extracted from our FCOM yielded good classification performance comparable to the other methods on the color texture.

Chiang Mai University Library

Address: CHIANG MAI

Email: cmulibref@cmu.ac.th

Name: Prabhas Chongstitvatana

Role: chairman

Name: Nipon Theera-Umpon

Role: Member

Name: Karn Patanukhom

Role: Member

Name: Sansdnee Auephanwiriyakul

Role: Member

Name: Patiwet Wuttisamwattana

Role: Member

Created: 2016

Modified: 2018-11-04

Issued: 2018-11-04

วิทยานิพนธ์/Thesis

application/pdf

CallNumber: Th 629.89 Y95I

eng

DegreeName: Doctor of Philosophy

Level: Doctoral Degree

Descipline: Computer Engineering

Grantor: Chiang Mai University

©copyrights Chiang Mai University

ลำดับที่.	ชื่อแฟ้มข้อมูล	ขนาดแฟ้มข้อมูล	จำนวนเข้าถึง	วัน-เวลาเข้าถึงล่าสุด
1	enco71216ymk_full.pdf	20.56 MB	13	2026-05-26 18:16:01

ใช้เวลา

0.031058 วินาที