International Journal of Advanced Networking and Applications

Ebrahim Parcham ¹, Monireh Pournazari ², Mina Hojati ³, Mehrdad Jalili Monir , Bahareh Mirzaei ⁴

Affiliations
1 Electrical and Computer Engineering Department, Tehran Science and Research University, Tehran, IR
2 Department of Electrical Computer and Biomedical Engineering, Qazvin Branch Islamic Azad University, Qazvin, IR
3 Department of Electrical Computer and Biomedical Engineering, Qazvin Branch, Islamic Azad University, Qazvin, IR
4 Sohrevardi Private High Educational Institute, School of Computer Engineering, Qazvin, IR

Abstract

Classifying similar images is one of the most interesting and essential image processing operations. Presented methods have some disadvantages like: low accuracy in analysis step and low speed in feature extraction process. In this paper, a new method for image classification is proposed in which similarity weight is revised by means of information in related and unrelated images. Based on researchers' idea, most of real world similarity measurement systems are nonlinear. Thus, traditional linear methods are not capable of recognizing nonlinear relationship and correlation in such systems. Undoubtedly, Self Organizing Map neural networks are strongest networks for data mining and nonlinear analysis of sophisticated spaces purposes. In our proposed method, we obtain images with the most similarity measure by extracting features of our target image and comparing them with the features of other images. We took advantage of NLPCA algorithm for feature extraction which is a nonlinear algorithm that has the ability to recognize the smallest variations even in noisy images. Finally, we compare the run time and efficiency of our proposed method with previous proposed methods.

Keywords

Self Organizing Maps (SOM), Nonlinear Dimensionality Reduction, Recognizing Content-Based Images, Artificial Neural Networks, Feature Vector, Machine Learning, Support Vector Machine, Clustering.

Full Text