Bobkov A.V. - Image registration in the real time applications (779134), страница 27
Текст из файла (страница 27)
The idea is tofind an invariant for each pair of features that probably match each other, so thatinvariant will be free from distortions. The distortion parameters are determined byfeature pair analysis, and appropriate correction is done. After that the feature canbe accounted in the search space by the ordinal way. It is easy to see that using ofinvariant features is a modification of search space decomposition method andhence it ascends all of its negative properties. Nevertheless, discussion of invariantfeature using method is outside the bounds of this research and it is a theme ofseparate research.In such a way, the significant factor affected to the reliability of theproposed method is a presence of unaccounted geometric distortions. If theirparameters are known, these distortions must be compensated in the vectored scenedescription, since it is not required significant computations in comparison to othermethods.5.1.4.4.
Image type and contentsImage content is always a complex task when a wide range of images mustbe processed. The problem is that the image can contain too many objects ofinterest, which makes image analysis too complex for real time mode, or – viceversa – image can contain no objects, and image processing becomes useless. Theobjects on the image can also have slightly different quality. So the properselection of image representation becomes very important task. Representationmust allow dealing with objects that partially invisible, overlapped by other objectsor frame bounds, or have bad quality.Quality of extraction process of selected features is determined by usedvectorization algorithm.
For example, when the image must be represented as a setof line segments, the contour tracing approaches will better extract shorter lines.Longer lines, which probably have gaps or which are not exactly direct lines, willbe separated to shorter ones. And vice versa, HT-based vectorisation will highlightlonger lines regardless gaps and curvature, and will suppress shorter lines.
Sinceproposed vectorisation method uses HT approach, it will also have a problem whenfacing images that mainly consist of short line segments.125There are several important classes of aerospace scenes, which mainlyconsist of short lines. Firstly, there are town scenes of some scale: long objects(streets etc.) on the image are produced by short ones (like houses and buildings).They cannot be united in one long object due to non-coaxiality and significant gapsbetween them. Same problems appeared with images of complex coastal line.
Onemore complex kind of scene is large-scale images: they have no man-made straightobjects but there is many nature objects of complex shape instead (e.g. mountainchains).When images of such kind are prevailing, the proposed vectorisation methodrequires additional tuning or it even need a serious modification. For example, it ispossible to make the vectorisation algorithm to extract shorter lines that usuallysuppressed (like separate building walls in large-scale town scenes). This causessignificant increasing of line amount in the scene description. This large amount oflines will significantly slow down either line matching or line gluing. Such amodification will also slow down the processing of all other scenes that does notneed small lines for a reliable image registration.In such a way, one more factor that affected to the image registrationreliability is contents of image itself and quality of objects on it. Significantdeviation of object characteristics (e.g. prevailing of objects of small sizes orcomplex forms) can require re-tuning or modification of the proposed vectorisationalgorithm.5.2.
Comparison of image registration techniques5.2.1. Line segment matching vs control point matchingProposed method of line segment matching has significant advantages incomparison to popular and commonly used control point matching method. Linesegments are more regular and common features on the image, than control points.Algorithm of line segment detection, proposed in the chapter 3, is a more formaland more reliable, than control point detector.
A set of line segments can be easilystructured in order to separate features of different importance, that is hard to dofor control points – they carry less information. As a result, registration algorithmcan use exactly such amount of lines that allowed by a hardware performance. Inthe same case the control point matching algorithm would skip some pointsregardless their importance.Older algorithms of HT-based line matching [Stockman et al 1982, Davies1991] were extracting control points as points of line joining and interception, withrelated problems. Proposed algorithm allows avoiding this.126The line segment matching works slower, when equal quality of lines andcontrol points are used, since it requires additional computations.
Thecomputational complexity of both algorithms is the same.5.2.2. Line segment matching vs. graph matchingGraph matching, in comparison to the line segment matching, allows reachhigher accuracy and reliability. However, they also have higher computationalcomplexity, which makes harder to use them in real time applications.Both accuracy and reliability of line segment matching can be raised usingcluster analysis for the responses in parameter space, as it was proposed in[Stockman et al 1982].5.2.3. Line segment matching vs. area-based methodsArea-based method (least square exhaustive search, extreme correlation,phase correlation in frequency domain etc), in common, allows using allinformation about image and thus they are potentially more reliable.
However,they also requires significant amount of computations. Even if specific hardware isused (e.g. for Fast Fourier transform computation or optical correlation), the speedof area-based registration remains low in comparison to line segment matching.Another serious disadvantage is sensibility to the image brightness variations,whenever line matching is free from it. Another specific factor is that area-batedmethods cannot be modified to find complex transformations of the image.However, another factor was found that affects on line segment matchingreliability – presence of unaccounted geometric distortions.
These distortions areinsignificant factor in area-based registration since matching areas of both imageswill change insignificantly. However, object contours will change significantly andit can happen that it is impossible to match both contours using accounted group oftransformations.
Therefore, the feature-based methods must use geometriccorrections always if it is possible.Area-based methods are less sensitive to the high frequency noise, since theline segment matching uses derive operation that highlights such a noise. However,there is some kind of specific noise, which causes crush of area-based methods.The example is the presence of clouds on the aerial images.
Area-based methodsrequire the clouds to be detected and replaced by white noise for the correct imageregistration. The line segment matching does not require additional processingsince cloud images contain insignificant amount of ling and direct lines.1275.3. ConclusionThe primary advantage of the proposed method is its significantperformance that becomes especially sensitive when images of the large size areprocessed. Therefore, the method can be the best alternative in the applicationswhere high performance is the primary task.The accuracy of the method has the range of one pixel if no uncompensatedimage distortion is present.
This is enough for the most of usual; applications.However, specific problems arise due to significant memory requirementsand lower reliability on the some classes of images.The significant disadvantage of the method is its significant memoryrequirements. The typical approach to lower these requirements can be hashingaccumulator array, implementing iterative search and search space decomposition.First two approaches requires significant amount of additional computations,whenever last one lowers the probability that desired position is found correctly.The selection of one or another approach in a specific application will bedetermined by task conditions and requirements.The lower reliability of the image registration is another important limitationof the method.
The method is not sensitive to noise presence and brightnessvariation. However, two new factors appear that affected to reliability andaccuracy: influence of unaccounted geometric distortions and affections of theimage content.If the parameters of geometric distortions are known or can be estimated,they must be compensated in the vectored scene description, since it does notrequired significant computations in comparison to other methods. On the oppositecase, their presence causes significant reducing of reliability and accuracy.Variations of the image contents become more important factor when theproposed method is used.
A significant deviation of object characteristics (e.g.prevailing of objects of small sizes or complex forms) can require re-tuning ormodification of the proposed vectorisation algorithm. However, parameter tuningcan be performed automatically by using measured dependences between thresholdlevel, actual and desired number of lines, and primary characteristics ofperformance, accuracy and reliability.In such a way, the method of the image registration proposed in this researchcan be considered as a good alternative in real time applications.
The primary area,for which the method was developed, is the aerial and space orientation usingvisual images. However, the method can be successfully applied in other areas thatrequire the image registration in real time or over large images. However, themethod limitations must be kept in mind. They are related with loss of reliability ifunaccounted geometric distortions are present on the image or when image doesnot contain enough amounts of long direct lines required for uniform imagematching.128CONCLUSIONThis research is devoted to a task of image registration for real timeapplications.Existing image registration techniques were examined, and structure of thenew method was proposed.
The method for fast and reliable image registration canbe implemented using fast scene vectorisation and further line matching.Existing image vectorisation techniques was also examined in order to selectmethods for a new vectorisation approach. The method must use edge extractionand line detection.An edge detection method that combines enough speed, reliability andaccuracy was proposed and compared with existing techniques. Although thecomponents of this method are already known, the proper selection of componentsis a complex task. So it required additional analysis to construct proper edgedetection.A new approach – extended Hough transform – is proposed and used todevelop a novel line segment detection method. The method characteristics wereinvestigated. These characteristics provide significantly better ratio of speed,accuracy and reliability, than that is possible using known method.
So obtainedline detector was selected as a base for a next stage of image registration – linematching.Possible variants of line matching were analysed in order to find anapproach with better characteristics. It was found that method of line segmentmatching based on Standard Hough Transform could provide significant speedwhile keeping accuracy and reliability high, as it is possible for selected class ofthe methods.
![](https://s.studizba.com/z.php?f=/templates/si/i/noavatar.png&w=100&h=100&t=1"){kind=avatar}
