Performance analysis of classification algorithms applied to Caltech101 image database

摘要

Identifying the wide range of applications, machine learning algorithms proved its ability to learn without being explicitly programmed. Classifying the images through machine learning algorithms is getting wide range of acceptability nowadays. Being a branch of Artificial Intelligence, machine learning implies the study of systems which has the capability to learn from data. Machine learning involves two parts - representation and generalization. Representation implies labeling seen data instances and generalization determines whether the system can perform well on unlabelled data instances. In this article, we focused on the performance of machine learning algorithms [1]. A CBIR (Content Based Image Retrieval) frame work has been developed and obtained a reduced texture feature data set using Caltech101 image database [2]. We highlight the top five algorithms such as Logistic, Bagging, LMT, Multiclass classifier and Attribute selection classifier which can be used for image classification. In introduction, an overview of the selected techniques is presented. We have extracted 2037 feature vectors from Caltech101 image database. These data are used to distinguish the performance of machine learning algorithms. Having checked all machine learning algorithms supported, we identified top five algorithms that have a better performance compared to other machine learning algorithms. The software used for testing is WEKA [3], which is an open source software developed by University of Waikato, New Zealand.