Bobkov A.V. - Image registration in the real time applications, страница 26

The typical approach to lower theserequirements can be hashing accumulator array, implementing iterative search andsearch space decomposition. First two approaches requires significant amount ofadditional computations, whenever last one lowers the probability that desiredposition is found correctly. The selection of one or another approach in a specificapplication will be determined by task conditions and requirements. For instance,the hashing can be best choice for the task of merging separate air photos intocomplete map. This task will produce a huge search space but is not significantlysensitive to computation amount. Another typical task is image registration for thetasks of autonomous navigation.

The most popular and suitable strategy here ishierarchical search since it provides good trade-off between reliability andperformance. Many machine vision applications like object localisation andinspection can also use search space decomposition, since the light conditions andbackground can be amplified in order to provide correct line extraction and imagematching.5.1.3.

AccuracyModern theory of image registration counts the HT approach as havinglesser accuracy (several pixels) than other method can provide. So experimentswere done in order to analyse thoroughly an accuracy of proposed approach.Experiments, described in Chapter 4, demonstrate that accuracy dependsmainly on amount of lines involved in matching process. When amount of lines aresmall (20 pixels and lesser), the standard deviation became greater than one, so theerror of several pixel can be obtained. This situation is typical for the objectrecognition applications, where relatively simple object can be described by smallamount of lines.

The inaccuracy value depends also on the prevailing lineorientation: as a rule, prevailing vertical and horizontal lines provides betteraccuracy in comparison to lines of other orientation.The situation significantly changes, when larger amounts of lines are used.In this case the standard deviation becomes lesser 0.01, and accuracy is in the onepixel range. The prevailing line orientation does not significantly affected toaccuracy, because statistical properties of line distribution compensate most ofother factors.There are several ways that can help to achieve even sub-pixel accuracy. Itbecomes possible since more complex object (lines) is used for matching instead ofseparate key points. One of approaches that allow obtaining sub-pixel accuracyrelated with an analysis of cluster shape in the search space.

In the simplest case itis enough to calculate a mass centre of the cluster. However, in some cases moresophisticated analysis is required, that takes into account moments of higher level.122The task of obtaining sub-pixel accuracy lies beyond the ranges of thisresearch, and this task is not reviewed here: sub-pixel accuracy can be more easilyobtained by some other methods like graph relaxation when approximate positionis already estimated by proposed method.The accuracy that can be achieved is sensitive to anotherfactor – uncompensated geometric distortion of one image to be registered. Whensuch distortions are present, the uniform matching becomes impossible in the termsof line sets.

It is impossible to propose such a translation that causes exact linematching on both images. In this case the response distribution in the parameterspace becomes significantly different from normal (it have a shape of deformedrings), and cannot be analysed by standard approaches. More complex (and morecomputationally expensive) methods are required in this case.So, the proposed method has accuracy in the range of one pixel if nouncompensated image distortion is present.5.1.4. ReliabilityAs a rule, feature based method has lesser reliability in comparison to areabased ones – this is a price for its significant speed.

The reliability depends onmany factors, and most important of them are the follows: changing of brightnessand contrast, high spatial frequency noise, geometric distortion and image content.5.1.4.1. Brightness and contrast variationsThe changing of brightness and contrast causes the difference in thebrightness value of appropriate pixel of both images, even the images are matchedexactly.

In in-door scenes it can be compensated by proper selection of lightconditions whenever it is impossible to do this in out-door scenes due to dailychanges of light conditions. The more complex problems rise in airspace imageregistration: additional brightness changing is caused by season and weathervariations. Furthermore, the colour of local areas can change, so the brightnesstransformation becomes non-global, and cannot be characterised by only twoparameters of brightness and contrast.Brightness changing is a serious problem for area-based methods, and oftenit is one of the main reasons to use feature-based method instead.

Feature-basedmethods (and proposed method as well) are practically insensitive to the brightnessand contrast changing – neither global nor local – until it allows extracting ofappropriate features and does not corrupt image (e.g. like solarization does).5.1.4.2. High spatial frequency noiseA noise on the image, as a rule, has significantly high-frequency nature. Thismakes possible to implement noise filtration, since objects on the image havelower frequencies. Noise, as a rule, appears on the image reception and it is123specific for different types of receiving devices. The sample of specific noise canbe a speckle noise in infrared images.The high-frequency noise can make more difficult (or even impossible) thefeature extraction, because the feature extraction process requires the edge pixelextraction, and edge pixels itself are the high-frequency object.

So when filtrationis used, edges can be suppressed or distorted as well: corners become rounded, tinydetails are disappeared, lines are shifted from its real position etc.The using of direct lines in comparison to other possible features (corners,reference points, arcs etc.) is more useful when significant high-frequency noise ispresent, since line detection is less sensitive to such a noise (lines are more lowfrequency objects).

In common case, the larger and more complex object is used,the less its extraction sensitive to noise. So, for example, it will be better to userectangles instead of separate line segments, line segments – better than referencepoints and arcs, and so on. However, the more complex is used, the rare it will bepresent on the image completely or even partially. So the line segments seem to bea good compromise between stability to noise and rarity on the image.If the static filtration with constant parameters is used, the problem of theline shifting has a simple solution: it is enough to implement the same filtration tothe both images. This causes the same line shifting on the both images.

It is not aproblem that lines will be shifted from its real position – it is important that bothimages will match exactly in their line set representation form.In such a way, presence of significant high-frequency noise makes moredifficult to extract object contours and further line extraction, and use of filtrationcauses line shifting and reduce both reliability and accuracy of the imageregistration.5.1.4.3. Geometric distortionsGeometric distortions are such a spatial transformation, which parametersare not determined in the image registration process. A presence of uncountedgeometric distortions causes the images cannot be matched exactly usingaccounted transformations.

For example, when one image is rotated in relation toanother, it is impossible to achieve image matching by implementing shifts only.Like the high-frequency noise, the geometric distortions played insignificantrole when area-based methods were used: regardless of geometric distortions andnoise, the most of image pixels will have matching pair.

The situation changeswhen feature-based methods are used, since amount of features in the imagedescription is significantly lesser (from 100 to 1000 times) than amount of pixelson it. As it was showed in the experiments in the Chapter 4, presence of eveninsignificant geometric distortions causes serious reducing of reliability of theimage registration process.

As a result, all geometric distortions must be accountedas full as it possible.If the parameters of geometric distortions are known – a priory or obtainedfrom other sources – their compensation is relatively simple. It is enough toimplement needed transformations over the vector description of the image.

Forexample, when the image is presented as a set of line segments, the124transformations must be applied to line segment end co-ordinates. The amount ofline segments significantly small (in comparison to overall amount of pixels thatwould be corrected if the area-based methods are used instead), so amount ofcomputations required for performing a correction is also small (it takes lesser than1% of total computation time).Such a correction can be implemented, for example, to a projectivedistortions in the case when camera is looks to a scene under an angle or whenscene is not flat, or when non-linear distortions caused by wide-angle camerausage are present.

Correction can be also implemented to rotations and scalingwhen their parameters are estimated from other sources.The problems arise when information about distortions is not available. Inthis case the task complexity is significantly arises, and acceptable solutions arenot known for such kind of tasks.One of the possible approaches is using of invariant features.