New Neural Buildings Stereo Matching Method Applied to Very High Resolution Ikonos Images

New Neural Buildings Stereo Matching Method Applied to Very High Resolution Ikonos Images

E. Zigh (National Institute of Telecommunications and Information Technologies and Communications of Oran, Algeria)
DOI: 10.4018/978-1-4666-7258-1.ch010
OnDemand PDF Download:
No Current Special Offers


The author introduces a new neural stereo matching method using very high resolution IKONOS images. They do not have the parameters of the images acquisition system or other technological resources like digital elevation model, Lidar, or Laser data. These images contain dense urban scenes including various kinds of roads, cars, vegetation, and builds. The author is interested by buildings; they have different shapes, positions, and intensity levels or colours, so they make a lot of “false matches.” To solve this issue, the authors extracts regions of buildings at first; after that, she proposes a neural stereo matching method. A neural field is chosen due to its good management of imprecision and uncertainty relatives to real problems in general and to this one in particular. To show the effectiveness of a proposed method, the chapter contains at first details about encountered problems, and secondly, it explains the stereo matching process, its different kinds, and a chosen approach; thirdly, it gives obtained results using panchromatic and colour images.
Chapter Preview


Stereo matching consists in finding homologous primitives using two images of the same scene. It is a crucial step in many applications like in telerobotic to assist a human operator in judging distances, marking landmarks for localization purposes, and identifying desired objects in the environment, also in medical images registration to compare or integrate the data obtained from different measurements and in remote sensing field to construct three-dimensional models very useful to monitor cities evolution, to manage risks (floods and earthquakes) and to improve telecommunication networks.

The construction of three-dimensional models needs generally four main steps: calibration, primitives extraction, stereo matching and reconstruction.

The result of each step influences on the next one, the goals of this chapter are: firstly, to deal with primitives extraction step and stereo matching processes, we operate on stereo panchromatic and stereo colour remote sensing images. These images are characterized by a very high spatial resolution equal to one meter, it means that each image pixel represents a land surface of 1 meter by 1 meter.

Also, the pairs of images are obtained from Ikonos satellite, they are delivered alone without parameters of the images acquisition system or another technological resources like digital elevation model, Lidar or Laser data.

The second goal of this chapter is to achieve reliability, automaticity, and low solution cost.

For that, we have proposed two methodologies, the first one is applied on panchromatic images, it consists of primitives extraction step including three stages: pre-processing, split-merge segmentation technique and thresholding. A pre-processing stage allows us an equitable illumination of a pair of images in order to make easier a next treatment, it is done using an histogram modification technique. Concerning segmentation step, we know that it is always difficult, even impossible to segment in a bi-univocal way two images of the same scene in order to match them, it does not exist an universal segmentation method which can apply to any image successfully (Chehata, 2006), so, we have tried to choose a method that can be adapted to our type of images and considered to be “satisfactory” if it realized a “good” compromise between algorithmic cost and segmentation quality. It is the Split-Merge method which consists in dividing the initial image into quadtree, after, we split or merge these regions according to logical predicates including some criteria like the grayscale homogeneity. We have obtained from this step a segmentation of all primitives of image, nevertheless, we are interested only by buildings primitives, for that, we have applied at next a thresholding technique.

After that, to find automatically correspondence buildings, a neural stereo matching technique is proposed, it is based on new constraints. Generally, previous stereo matching researches used some known constraints like epi-polar, fronto parallel, ordering and uniqueness constraints, however, satisfying all these constraints in our stereo images isn’t a simple task, because they cover urban dense scenes with variable inclined buildings from right to left image covering the same scene, so, we couldn’t often check ordering constraint for example, also, epi-polar constraint couldn’t be checked because we haven’t had the parameters of images acquisition system.

The originality of a proposed neural stereo matching technique is the use of new constraints having two main advantages: a good description of extracted primitives and a simplicity of implementation, there are

surface, elongation and gravity centers coordinates as new geometrical constraints and an average of intensity as new photometric constraint. A proposed neural network is a Hopfield type, it solves the optimization problem by minimizing a cost function whose minimum value represents the best solution, the nodes are the assumptions (the possible correspondences) and the connections between them are the constraints.

Key Terms in this Chapter

Stereo Matching: It deals with finding point correspondences among two views of a scene. This is an important step towards a number of recognition algorithms; e.g., 3D modeling. A reliable matching algorithm results in a good 3D estimation of object structures. It is also called stereo correspondence and it can be broadly classified into two categories: the intensity-based matching and the feature-based matching techniques. In the first category, the matching process is applied directly to the intensity profiles of the two images, while in the second, features are first extracted from the images and the matching process is applied to the features.

Radiometric Resolution: It determines how finely a system can represent or distinguish differences of intensity, and is usually expressed as a number of levels or a number of bits, for example 8 bits or from(0-255) levels that is typical of computer image files. The higher the radiometric resolution, the better subtle differences of intensity of reflectivity can be represented, at least in theory. In practice, the effective radiometric resolution is typically limited by the noise level, rather than by the number of bits of representation.

Spatial Resolution: The area on the earth’s surface that can be seen by a sensor as being separate from its surroundings and is represented by a pixel. It is the projection of a detector element or a slit into the ground. In other words scanners spatial resolution is the ground segment sensed at any instant. It is also called ground resolution element. The spatial resolution at which data are acquired has two effects –the ability to identify various features and quantify their extent. Higher is a spatial resolution, better is a detail of image elements.

Remote Sensing: In the broadest sense, remote sensing is the measurement or acquisition of information of an object or phenomenon, by a recording device that is not in physical or intimate contact with the object. In practice, remote sensing is the utilization at a distance (as from aircraft, spacecraft, satellite, or ship) of any device for gathering information about the environment. Remote sensing can either be passive or active. Active systems have their own source of energy whereas the passive depend upon the solar illumination. The used images in this research are obtained from a passive acquisition.

Spectral Resolution: The range of wavelength that satellite imaging system can detect, it refers to the width and number of spectral bands. The narrow band, the greater spectral resolution.

Ikonos 2: It is a commercial earth observation satellite which was launched in 1999, and it was the first to collect publicly available high-resolution imagery at one and four meter resolution. It offers multispectral (MS) and panchromatic (PAN) imagery. The images obtained from a satellite have some characteristics like: spatial resolution, spectral resolution, temporal resolution and radiometric resolution. It describes the ability of the sensor to define fine wavelength intervals i.e. sampling the spatially segmented image in different spectral intervals, thereby allowing the spectral irradiance of the image to be determined.

Temporal Resolution: It is a measure of the repeat cycle or frequency with which a sensor revisits the same part of the Earth’s surface. The frequency will vary from several times per day, for a typical weather satellite, to 8—20 times a year for a moderate ground resolution satellite, such as Landsat TM. The frequency characteristics will be determined by the design of the satellite sensor and its orbit pattern.

VHR: Very high resolution.

Complete Chapter List

Search this Book: