Mixed Pooling for Remote Sensing Image Feature Extraction and Retrieval

Abstract

Advancement in satellite technology has generated huge database of remote sensing (RS) image.  From this image database retrieving required data is challenging. Thus, current supervised content based image retrieval systems are not competent to characterize and exploit the high level meaningful content of RS images for retrieval problems. To overcome these problems with the existing method this article deals with key aspects of Remote sensing retrieval and provides a detailed overview of current methods. The pre-trained CNN method is used with ImageNet dataset to extricate the attributes of high resolution remote sensing images.  The VGG 16, VGG19 and GoogLeNet CNN architectures are used to retrieve remote sensing images. The retrieval performance is measured at different layers of CNN along with three different pooling methods. The simulation result is carried out on publicly available dataset UCM and WHU-RS dataset.  The proposed technique gives improved performance accuracy compared to existing standard CBIR methods.