Performance Evaluation of Image Segmentation and Texture Extraction Methods in Scene Analysis

The main aim of this thesis is to evaluate the performance of image segmentation and texture analysis algorithms on synthetic and real images. As a part of this study, two popular texture benchmarks called MeasTex and VisTex have been used. A new scene analysis benchmark, called PANN database, has been generated as a part of this study for the evaluation of image analysis tools on natural object recognition tasks. The thesis demonstrates the considerable variability in an image understanding system performance based on different choices of image segmentation and texture analysis algorithms used. Hence, it is proposed that optimising each image analysis tool in a chain of processes is not enough for deriving optimal performances. The effect of a preceding process on its successor is both important and significant. So for example, how well we segment images, can have a dramatic impact on the quality of texture features we extract from such regions and subsequently this impacts our object recognition ability. This thesis includes results of exhaustive experimentation with four image segmentation algorithms, five texture analysis algorithms, and two classifiers, to demonstrate this variability. The thesis also discusses the similarity and differences in performances highlighting different patterns of mistakes made by different combinations. In addition to being a comparative study, it also shows results on object recognition in natural scene images. A complete system starting from image acquisition to generating ground truth data and classifying it is described for the analysis of natural scenes.