An Intelligent Multi-Resolutional and Rotational Invariant Texture Descriptor for Image Retrieval Systems

Abstract

To find out the identical or comparable images from the large rotated databases with higher retrieval accuracy and lesser time is the challenging task in Content based Image Retrieval systems (CBIR). Considering this problem, an intelligent and efficient technique is proposed for texture based images. In this method, firstly a new joint feature vector is created which inherits the properties of Local binary pattern (LBP) which has steadiness regarding changes in illumination and rotation and discrete wavelet transform (DWT) which is multi-resolutional and multi-oriented along with higher directionality. Secondly, after the creation of hybrid feature vector, to increase the accuracy of the system, classifiers are employed on the combination of LBP and DWT. The performance of two machine learning classifiers is proposed here which are Support Vector Machine (SVM) and Extreme learning machine (ELM). Both proposed methods P1 (LBP+DWT+SVM) and P2 (LBP+DWT+ELM) are tested on rotated Brodatz dataset consisting of 1456 texture images and MIT VisTex dataset of 640 images. In both experiments the results of both the proposed methods are much better than simple combination of DWT +LBP and much other state of art methods in terms of precision and accuracy when different number of images is retrieved.  But the results obtained by ELM algorithm shows some more improvement than SVM. Such as when top 25 images are retrieved then in case of Brodatz database the precision is up to 94% and for MIT VisTex database its value is up to 96% with ELM classifier which is very much superior to other existing texture retrieval methods.