Computation Performance Gain in Key Feature Extraction for CBIR system for Remote Sensing Images

Abstract

Remote sensing data is increasing significantly fast and creating requirement for an e cient CBIR system. All content based image retrieval system starts with feature extraction or description which are expected to be exhaustive, capturing all essential contents of image. Remote Sensing data are acquired in large number and hence added continuously in the database. This paper proposes to deploy GPU to obtain the computation gain for the feature extraction. Morphological descriptors like CCH and RIT and color descriptors like LBP represents the texture information of the image are well established in [1, 2] but are extremely compute intensive. The Parallel implementation of the above mentioned features are proposed and are experimented on UC Merced LULC data set. The experimentation results show 380X speedup for LBP, 20X in case of CCH and RIT which is a significant gain. This enables the availability of the added images in the dataset almost in real time, for matching and retrieval.