Development of a Framework for Stereo Image Retrieval With Both Height and Planar Featuresby Feifei Peng, Le Wang, Jianya Gong, Huayi Wu

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing


Computers in Earth Sciences / Atmospheric Science


Image retrieval using both color and texture features

Dong-cheng SHI, Lan XU, Ling-yan HAN

Stereo effect of image converted from planar

Ran Liu, Qingsheng Zhu, Xiaoyan Xu, Liou Zhi, Hongtao Xie, Jun Yang, Xiaoyun Zhang

The American Academy of Fine Arts in Rome

Adapted for Brush and Pencil

The application of synthetic features in image retrieval

lihong Li, Xingyu Gong, Yuanfei Xu



Development of a Framework for Stereo Image

Retrieval With Both Height and Planar Features

Feifei Peng, Le Wang, Jianya Gong, and Huayi Wu

Abstract—The wide availability and increasing number of applications for high-resolution optical satellite stereo images (HrosSIs) have created a surging demand for the development of effective content-based image retrieval methods. However, this is a challenge for existing stereo image retrieval methods since they were designed for stereo images collected from closerange imaging sensors. Thus, successful retrieval of images is not assured given the mismatch between existing methods and the characteristics of HrosSIs. Moreover, none of the existing remote sensing image retrieval methods takes account of the specific characteristics of HrosSIs such as the viewing number and multiview angles. This paper proposes a generic framework to exploit the unique characteristics of HrosSIs data so as to allow efficient and accurate content-based HrosSI retrieval.

HrosSIs retrieval is executed by similarity matching between the features obtained from digital surface models (DSMs) and orthoimages, both extracted from the HrosSIs. In addition, the significance of height information for HrosSI retrieval was investigated. A prototype system was designed and implemented for method validation using the ISPRS stereo benchmark test dataset. Experimental results show that the proposed techniques are efficient for HrosSI retrieval. The proposed framework is efficient and suitable for spaceborne stereo images but might also be suitable for airborne stereo images as well. Experimental results also show that height information alone is inefficient and unstable for HrosSI retrieval; however, a combination of height information and planar information is efficient and stable.

Index Terms—Digital surface model (DSM), fractals, height features, image retrieval, orthoimage, planar features, stereo imagery.


HrosSI high-resolution optical satellite stereo image.

DSM digital surface model.

GSD ground sample distance.

D fractal dimension.

RFD regional fractal dimension.

RFI regional fractal image.

Manuscript received March 20, 2014; revised September 09, 2014; revised

June 19, 2014; accepted October 09, 2014. Date of publication November 12, 2014; date of current version February 09, 2015. This work was supported by the National Key Basic Research and Development Program of

China under Grant 2012CB719906. The work of L. Wang was supported in part by the National Science Foundation (NSF) under Grant DEB-0810933 and

Grant BCS-0822489 and in part by the State Key Laboratory of Information

Engineering in Surveying, Mapping and Remote Sensing (LIESMARS),

Wuhan University.

F. Peng, J. Gong, and H. Wu are with the State Key Laboratory of Information

Engineering in Surveying, Mapping and Remote Sensing (LIESMARS),

Wuhan University, Wuhan 430079, China (e-mail:;;

L. Wang is with the Department of Geography, The State University of New

York at Buffalo, Buffalo, NY 14261 USA (e-mail:

Color versions of one or more of the figures in this paper are available online at

Digital Object Identifier 10.1109/JSTARS.2014.2363953

NRFCM normalized regional fractal cooccurrence matrix.

NDRI normalized dissimilarity ranking index.

CS1 Cartosat-1.

WV1 Worldview-1.


M ANY SATELLITES with the capability to pro-duce high-resolution optical satellite stereo images (HrosSIs) are currently available, such as IKONOS, QuickBird,

WorldView-1/2, Cartosat-1/2, GeoEye-1, Pleiades-HR, ALOS (PRISM), and ZY3 [1]. As a result, large quantities of HrosSIs are now accessible to researchers and the general public.

HrosSIs are widely applied in various fields, such as digital surface model (DSM) production [2]–[4], building reconstruction [5], [6], change detection [7]–[9], and hazard assessment [10]. In some applications, desired HrosSIs must first be located based on semantic information or content-based similarities between HrosSIs. HrosSIs covering human settlements [11] in China, for instance, must first be geographically located to assess human settlement areas. It is very difficult to infer image content from image metadata (e.g., rational polynomial coefficients, geographic coverage, and acquisition time), thus the desired images cannot be found by using image metadata alone.

Content-based image retrieval expedites search and discovery among large quantities of images based on the content of those images; and therefore can meet the discovery requirements when locating these desired HrosSIs.

The unique characteristics of HrosSIs data, however, make it difficult to extract uniform features from various HrosSIs for content-based HrosSI retrieval. HrosSIs are acquired by diverse remote sensors under diverse acquisition conditions and vary significantly in many ways, such as in stereo acquisition modes [12], viewing number, viewing angles, convergence angle, base-to-height ratio, ground sample distance (GSD), radiometric resolution, and metadata structure. These unique characteristics of HrosSIs data are not exploited in existing content-based image retrieval methods; thus, their potential in feature extraction and retrieval is unknown.

The significance of height information needs to be investigated to obtain efficient and accurate HrosSI retrieval. Height information can easily be derived from stereo-extracted DSMs, one type of primary product created from HrosSIs data.

However, height information is not taken into consideration in existing image retrieval methods. Height information may be efficient to distinguish between land cover types that are difficult to distinguish with planar features alone. For instance, forest land on a plain and forest land in mountains both 1939-1404 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.