Comparison Of Image Classification Techniques Using Caltech 101 Dataset

• Main Authors: Mokhairi, Makhtar, Engku Fadzli Hasan, Syed Abdullah, Nur Shazwani, Kamarudin
• Format: Article
• Language: English
• Published: 2015
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://eprints.unisza.edu.my/5748/1/FH02-FIK-15-02472.jpg; http://eprints.unisza.edu.my/5748/

This paper presents the technique for the classification of single object images. First, this paper aims to introduce the efficient technique in order to classify single object image. Second, each single methods uses in order to propose the techniques were elaborated in this paper. It start from image segmentation, object identification, feature extraction, feature selection and classification. Finally, the best classifier that can provide the best results were identified. The efficiency of the proposed method is define by comparing the result of classification using two different datasets from author’s previous paper. The obligation for development of image classification has been improved due to remarkable growth in volume of images, as well as the widespread application in multiple fields. This paper explores the process of classifying images by the categories of object in the case of a large number of object categories. The use of a set of features to describe 2D shapes in low-level images has been proposed. The proposed technique aims a short and simple way to extract shape description before classifying the image. Using the Caltech 101 object recognition benchmark, classification was tested using four different classifiers; BayesNet, NaiveBayesUpdateable, Random Tree and IBk. Estimated accuracy was in the range from 58% to 99% (using 10-cross validation). By comparing with Amazon data, it is proved that the proposed model is more suitable for single object image. Amazon images give higher accuracy with the range from 80% to 99.48%.