Classification Model for Skin Lesion Image

The problem of image classification mostly focuses on feature extraction which depends on image data. However, the famous feature is extracted from images using Gray Level Co-occurrence matrix (GLCM) in order to cover all feature, which are range differently. The problem is less accuracy when it is inputted into classification model. This idea of paper is proposed for feature normalization 2 types called local-normalization and global-normalization from all feature extraction using GLCM in preprocessing of classification method. These feature values extracted from GLCM are transformed to proper normalization and given input classification model as Back Propagation in Multi-layer perceptron and Multi-Class support vector machine methods (polynomial and RBF) to compare these classification models. The skin disease image classification which occurs from skin lesions has divided into four classes: Tinea Corporis, Pityriasis Versicolor, Molluscum Contagiosum and Herpes Zoster. The experimental results are shown comparison between non-normalization and normalization within the same class called local-normalization and all classes called global-normalization. The accuracy of MLP with normalization by min-max normalization with local-normalization is highest to 92%. The methods of polynomial-SVM and RBF-SVM are given accuracy as 85% and 81% respectively. Whereas, the accuracy of classification model with non-normalization, and global-normalization are given average of accuracy as only 35% approximately.