Feature Fusion for Efficient Retrieval in CBIR Systems

Abstract

Image indexing and retrieval has a wide spectrum of promising applications, motivating the interest of researchers worldwide.There is an amazing growth in the amount of digital data(Images and videos) in recent years. There is lack of efficient algorithms for classifying and retrieving the satellite images in the field of content based image retrieval(CBIR) systems. Satellite image retrieval is the demand of the time, with increasing resolution in pixel, frequency and time. The rate of image acquisition has gone up as never before and hence that confirms to the need of CBIR system. Several feature sets being proposed with the capability to represent different aspects of the images such as color moment for color features, Circular covariance histogram and Rotation invariant point triplets for texture information etc. But the fact is all the features have varied scales and ranges. Hence, it becomes mandatory to combine them optimally in different weights. This work proposes and implements a Late Fusion technique to optimize on the retrieval results. Deep Learning is an emerging branch of machine learning, where hand tuned features by humans are replaced with an added learning layer in the network which automatically learns the features. Deep learning like human brains learns features step-by-step. The work demonstrates experimentations for image retrieval using the concepts of deep learning with convolutional neural network.