An Evaluation of Image Texture Descriptors and their Invariant Properties

Image processing applications include image classification, image segmentation, image synthesis and many others. Each such task depends on extracting an effective set of features to characterize the images, and texture analysis has proven to output some of the most valuable features. For this reason, image texture analysis has been an actively researched topic and numerous methods have been proposed, each of them having its advantages and limitations. In practical applications, it is impossible to ensure that all images have the same scale, rotation, viewpoint, etc., so texture analysis methods should ideally be invariant. This study inspects the most commonly used operators for extracting textural features, tests their accuracy in classifying the Kylberg texture dataset, and evaluates their invariant properties by the means of various synthetically transformed images. By conducting this analysis, we identified the shortcomings of the existing approaches, and will be able to address them in our future work by formulating some improvements to existing operators to increase their accuracy and to make them invariant to a larger set of transformations.