I.J. Information Technology and Computer Science, 2015, 10, 61-73 Published Online September 2015 in MECS (http://www.mecs-press.org/) DOI: 10.5815/ijitcs.2015.10.08 Copyright © 2015 MECS I.J. Information Technology and Computer Science, 2015, 10, 61-73 Content Based Image Recognition by Information Fusion with Multiview Features Rik Das Department of Information Technology, Xavier Institute of Social Service, Ranchi, Jharkhand, India E-mail: rikdas78@gmail.com Dr. Sudeep Thepade Department of Computer Engineering, Pimpri Chinchwad College of Engineering, Pune, India E-mail: sudeepthepade@gmail.com Saurav Ghosh A.K. Choudhury School of Information Technology, University of Calcutta, Kolkata, West Bengal, India E-mail: sauravghoshcu@gmail.com Abstract—Substantial research interest has been observed in the field of object recognition as a vital component for modern intelligent systems. Content based image classification and retrieval have been considered as two popular techniques for identifying the object of interest. Feature extraction has played the pivotal role towards successful implementation of the aforesaid techniques. The paper has presented two novel techniques of feature extraction from diverse image categories both in spatial domain and in frequency domain. The multi view features from the image categories were evaluated for classification and retrieval performances by means of a fusion based recognition architecture. The experimentation was carried out with four different popular public datasets. The proposed fusion framework has exhibited an average increase of 24.71% and 20.78% in precision rates for classification and retrieval respectively, when compared to state-of-the art techniques. The experimental findings were validated with a paired t test for statistical significance. Index Terms—Local Threshold, Partial DCT coefficient, KNN Classifier, Fusion based Recognition, t test. I. INTRODUCTION Content based feature extraction has become an active area of research in the last few decades with proliferation of digital images from various image capturing sources and the internet [1, 40, 41]. Extraction of meaningful information from image content has been well perceived in spatial domain and in frequency domain [2]. Spatial domain techniques have considered binarization of images as a popular method for feature extraction [3, 4, 42, 43, 44, 45]. In the frequency domain, discrete cosine transform has proved to be an efficient tool to extract features from transformed images [5, 6]. Partial coefficient selection from transformed images has shown considerable improvement in retrieval results in the past [7]. Consequently, the process of content based feature extraction from images can be aimed for better description of the input data to facilitate image classification and retrieval with lesser computation load [8]. However, it has been observed that traditional feature extraction algorithms take out large size of feature set which scales up the space complexity and increases the training time [9, 10, 11]. In this paper, the authors have proposed two novel feature extraction techniques namely feature extraction by image binarization by applying Sauvola's local threshold selection technique and feature extraction by customized Discrete Cosine Transform (DCT). At the outset, the proposed techniques of feature extraction were evaluated for classification performances with Wang dataset (1000 images), Corel dataset (10,800 images), Caltech dataset (8,127 images) and OT Scene dataset (2,688 images). The authors have further proposed a fusion framework for content based image classification which has outclassed the existing techniques and has shown statistical significance in boosting up the content based image identification process. Subsequently, a retrieval process with focused query has been formulated which has outperformed the contemporary techniques of content based image retrieval. II. RELATED WORK Existing methods for feature extraction using binarization have been divided into three different techniques of threshold selection namely mean threshold selection, local threshold selection and global threshold selection. Extraction of features from bit planes and even and odd image varieties has been performed by mean threshold selection for better classification results [12, 13]. Feature extraction using ternary mean threshold [14] and multilevel mean threshold [15] for binarization has also considered mean value of the grey levels for selecting the threshold for binarization. But the spread of data was not considered for threshold selection process in