BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:On Gray Scale Features-Based Image Classification of Textural Type
  Continuous Objects - Andrei Reztsov\, Australian Key Centre for Microscop
 y &amp\; Microanalysis\, Sydney\, Australia
DTSTART:20060515T140000Z
DTEND:20060515T150000Z
UID:TALK4939@talks.cam.ac.uk
CONTACT:Edmund Ward
DESCRIPTION:A novel approach to Gray Scale Features-based Image Classifica
 tion is presented. Used by us classifier is Support Vector Machine (SVM). 
 Images were obtained for unbounded non-structured textural type continuous
  3D objects without contour lines or layers. These were cylindrical sample
 s (cores) drilled from amorphous porous carbon material\, to be sorted int
 o seven classes based on their structure. Their 3D X-ray microtomography (
 XRMT) data were studied. Although this method can be applied to a wide ran
 ge of 2D and 3D raster and vector images\, we present here only results re
 lated to XRMT images. The calculation of Gray Scale Features was done in a
  way similar to (Ronneberger et al.\, 2002)\, but we had to deal with almo
 st completely chaotic textures without any noticeable regularity inside or
  well defined boundaries that could be used as key registration points for
  Pattern Recognition or Image Classification. These principal differences 
 between data sets and 3D images (Ronneberger et al.\, 2002) led us to deve
 lop several pioneering ideas and techniques. See also (Schulz-Mirbach\, 19
 95). Seven types of mineral carbon material were studied. It was found tha
 t it is very difficult to classify them by using traditional image analysi
 s tools because of their complex natural structure and the fact that even 
 different samples of the same material from the same source often exhibit 
 significant differences in structural texture. More conservative routine o
 f visual categorization by human operator is labour consuming and is not f
 ully error-proof. Computational procedure was established to solve this pr
 oblem. Firstly\, we randomly selected 10 cylindrical cores of material in 
 addition to a statistically proven random way of positioning drilling. Dat
 a sets of size 1024x1024x1024\, obtained for these 10 cores\, were later u
 sed as original 3D images. The structure of each core differed from the ot
 her cores and there was significant variation of porosity and wall topolog
 y inside a particular core. To cope with this we used a statistically viab
 le procedure of defining approximately 100 sub-volumes each of 256x256x256
  size. This sub- sampling acquired the role played before by boundary and 
 completely black background (Ronneberger et al.\, 2002). Inside each sub-v
 olume we built a mesh of randomly distributed sets of concentric spherical
  shells of 12 radii. We then performed Monte Carlo numerical integration o
 ver each sphere using 3 functions based on the Gray Values themselves and 
 their local combinatorial relationships similar to (Ronneberger et al.\, 2
 002). For each texture type 240 vectors of Invariant Gray Scale Features w
 ere obtained and then used as input data for SVM statistical classificatio
 n. The corresponding numerical output showed exceptionally good recognitio
 n and classification results with successful recognition in more than 97% 
 of case. Further modification of the standard classification procedure is 
 suggested that makes close to 100% recognition possible.\n\nAs part of rec
 ommended data scaling for SVM analysis we computed multidimensional Mean a
 nd Standard Deviation Vectors and used them later for the computational st
 ability characterisation\, since it is impossible to obtain stability char
 acterisation analytically. We applied some controlled noise\, fluctuation 
 to Gray Scale Feature data and verified how much the SVM output changed. T
 he shape of the obtained stability curve and computed gradient in the poin
 t of original state (i.e. with no noise applied) characterised stability i
 tself\, SVM and porous material. See (Reztsov & Jones\, 2006\; Jones et al
 .\, 2006) for other methods of utilising Invariant Gray Scale Features for
  textural type continuous objects\, different internal properties of the d
 ata domain and the use of out technique for 2D density distributions of am
 orphous materials and for analysis of vector form data as used in atom pro
 be style instrumentation.\n\nReferences:\n\nSchulz-Mirbach\, H. (1995). In
 variant features for gray scale images. In 17 DAGM - Symposium "Mustererke
 nnung''\,  Bielefeld\, 1995. Sagerer\, G.\, Posch\, S. & Kummert\, F. (Eds
 )\, pp. 1-14. Reihe Informatik aktuell\, Springer.\n\nRonneberger\, O.\, S
 chultz\, E.\, & Burkhardt\, H. (2002).  Automated Pollen Recognition using
  3D Volume Images from Fluorescence Microscopy. Aerobiologia 18\, 107 115.
   \n\nReztsov A.\,V.  & Jones A.S. (2006). Invariant Gray Scale Features f
 or Pattern Recognition and Image Classification of textural type continuou
 s objects. In Australian Conference in Microscopy and Microanalysis\, Sydn
 ey\, 5th-9th February 2006\, p. 64.  \n\nJones A.S.\, Reztsov A.V. & Loo C
 .E. (2006). Application of Invariant Grey Scale Features for Analysis of P
 orous Minerals. Micron 38\, to appear. \n
LOCATION:T001 [Tower Seminar Room]\, Materials Science and Metallurgy\, De
 partment of
END:VEVENT
END:VCALENDAR