A systematic evaluation of the scale invariance of

Authors: Andreas Uhl Georg Wimmer

Publish Date: 2014/12/09

Volume: 18, Issue: 4, Pages: 945-969

PDF Link

Abstract

A large variety of wellknown scaleinvariant texture recognition methods is tested with respect to their scale invariance The scale invariance of these methods is estimated by comparing the results of two test setups In the first test setup the images of the training and evaluation set are acquired under same scale conditions and in the second test setup the images in the evaluation set are gathered under different scale conditions than those of the training set For the first test setup scale invariance is not needed whereas for the second test setup scale invariance is obviously crucial The difference between the results of these two test setups indicates the scale invariance of a method the higher the scale invariance the lower the difference The scale invariance of the methods is additionally estimated by analyzing the similarity of the feature vectors of images and their scaled versions Additionally to the scale invariance we also test eventual viewpoint and illumination invariance of the methods As texture databases for our tests we use the KTHTIPS database and the CUReT database Results imply that many of the considered methods are not as scaleinvariant as expectedTexture analysis is one of the fundamental issues in image processing and pattern recognition Techniques for the analysis of texture in digital images are essential to a range of applications in areas as diverse as robotics defence medicine and geosciences 27The majority of existing texture analysis methods works with the assumption that texture images are acquired from the same viewpoint 42 This limitation could make these methods useless for applications where textures occur with different scales orientations 2 or translations Surveys about existing scale and orientation invariant texture analysis approaches are found in 32 42 Scale invariance is also needed in other computer vision applications like eg image annotation 17 33 object recognition 18 medical image analysis 12 et ceteraMost of the scale invariant texture analysis approaches are tested on public databases like the Brodatz 4 the CUReT 5 the KTHTIPS 11 or the UIUCTex 14 database The scale invariance of these methods is founded on theoretical concepts but the question is if these methods do actually exhibit scale invariance in practice Most approaches are never really tested with respect to their effective scale invariance If some techniques provide good results for texture databases where textures occur at different scales then these methods are commonly assumed to be de facto scaleinvariantThe standard setup for testing approaches on texture databases is to construct an evaluation and a training set where the training set consists of a number of randomly chosen texture samples per texture class and the evaluation set of the remaining texture samples If the texture database consists of texture images at different scales and if the results of a method are good for this standard setup does that imply that the considered approach is scaleinvariant Not necessarily Especially if the training set consists of a higher number of texture images per class for nearly each image of the evaluation set there might be images of the same class in the training set with rather similar scales This means that a technique does not necessarily have to be scaleinvariant to work well on a texture database containing textures with various scalesAdditionally methods which are not scaleinvariant may provide good results only if they are able to extract important scale dependent information to differentiate between textures of various classes So feature expressiveness might dominate the issue of scale invariance

Keywords: