BOOSTING THE PERFORMANCE OF THE STATE-OF-ART IMAGE CLASSIFICATION TECHNIQUES

Abstract

In the modern times, Query by Image is the fad. With the Digital Image Capture Devices made available at very low cost, the image database has been expanding very rapidly. Challenge of maintaining such a humongous Image Database, searching images from it efficiently, is becoming more and more cumbersome. A well organised database can prove to be a great relief. This paper deals with the initial step of classifying images in a very efficient way to lead to a well organised image database from where query image could be retrieved swiftly. Many state-of-the-art Image Classification Techniques including Harris Corner Detection (HCD), Canny’s Edge Detection (CED), Laplacian of Gaussian (LoG), Scale Invariant Feature Transform (SIFT) and Speeded up robust features (SURF) are compared and contrasted. Various algorithms for key point matching viz. Oriented FAST Rotaed BRIEF (ORB), KAZE and Accelerated KAZE etc. neural and models like k-Medoids Clustering, k Nearest Neighbourhood (kNN) and Convolution Neural Network (CNN) etc. and  networks having various designs – LeNet, AlexNet, SqueezeNet, GoogLeNet, ResNet  etc. are compared and contrasted on the basis of their performances keeping in view various parameters like Precision, Recall and Speed. An improvisation is proposed to boost the performance of this phase on the way to a complete algorithm for content-based image retrieval system.