10 likes | 135 Views
Technique A. IMAGE. Quantified Characteristics Response variables. Technique B. Technique C. Linear Model. Analysis of Variance ANOVA. Optimizing Texture Feature Extraction in Image Analysis by Using Experimental Design Theory S. A. Orjuela Vargas 1 , R. de Keyser 2 , and W. Philips 1.
E N D
Technique A IMAGE Quantified Characteristics Response variables Technique B Technique C Linear Model Analysis of Variance ANOVA Optimizing Texture Feature Extraction in Image Analysis by Using Experimental Design Theory S. A. Orjuela Vargas1, R. de Keyser2, and W. Philips1 1.Department of Telecommunications and Information Processing (TELIN-IPI-IBBT), Gent University Sint-Pietersnieuwstraat 41, B-9000 Gent, Belgium. 2. Department of Electrical Energy, Systems and Automation (EeSA) , Gent University, Belgium http://telin.ugent.be/~seraleov seraleov@telin.ugent.be Texture Analysis Texture is a pattern describing variationsin a surface at scales smaller than the scale of interest To extract texture features several techniques are evaluated Optimal set of features are commonly chosen by “best guess approach” Coocurrence Matrices Industrial inspection medical imaging Laws’ Energy Local Binary Patterns Techniques dependences are not evaluated Wavelets Remote sensing Image retrieval No guarantee of optimality No reliability of results and inferences Texture analysis has four major areas of interest High amount of data must be analyzed Experimental Design Theory in Image Analysis Significant differences are identified when probability results are smaller than a given Characteristics from image outcomes are quantified in response variables to compare techniques A test for multiple comparison must be further performed = Comparison is performed by using ANalysis Of VAriance (ANOVA) Technique dependences are similarly detected Optimal combinations of techniques are found Application: Automatic wear label assessment for carpets, comparison of three techniques Carpets are certified according to their capability in retaining the original appearance The relationship between wear labels and features must be at least linear-ranked Significance differences detected for both response variables at 99% of confidence Two response variables are quantified Worn samples are compared to original samples assessing the change in appearance with wear labels • Monotonicity (): • Based on the Spearman’s rank correlation • 2. Number of consecutive wear labels that can be statistically distinguished (): • Based on the Tukey test Wear Labels Features LBPRMC technique performs better than the others for both response variables Distinction of different degrees of wear by texture analysis in photographs If so, the distinction between consecutive wear labels must be maximized A planning phase using experimental design theory add reliability to image analysis results [2] S. A. Orjuela, E. Vansteenkiste, F. Rooms, S. De Meulemeester, R. De Keyser, and W. Philips. Evaluationof the wear label description in carpets by using local binary pattern techniques. Textile Research Journal, 2010. [1] S. A. Orjuela, Experimetal Design Theory as a Tool for Optimizing Feature Extraction in Image Analysis, Tutorial, STSIVA 2010. 11th FIRW PhD symposium | December 2010