Newell, AJ; Griffin, LD; Morgan, RM; Bull, PA; (2011) Texture-based estimation of physical characteristics of sand grains. In: Proceedings: 2010 Digital Image Computing: Techniques and Applications: DICTA 2010 1-3 December 2010, Sydney, Australia. (pp. pp. 504-509). IEEE Computer Society: Piscataway, US.

Preview

Text 1308978.pdf Available under License : See the attached licence file. Download (411kB)

Abstract

The common occurrence and transportability of quartz sand grains make them useful for forensic analysis, providing that grains can be accurately and consistently designated into prespecified types. Recent advances in the analysis of surface texture features found in scanning electron microscopy images of such grains have advanced this process. However, this requires expert knowledge that is not only time intensive, but also rare, meaning that automation is a highly attractive prospect if it were possible to achieve good levels of performance. Basic Image Feature Columns (BIF Columns), which use local symmetry type to produce a highly invariant yet distinctive encoding, have shown leading performance in standard texture recognition tasks used in computer vision. However, the system has not previously been tested on a real world problem. Here we demonstrate that the BIF Column system offers a simple yet effective solution to grain classification using surface texture. In a two class problem, where human level performance is expected to be perfect, the system classifies all but one grain from a sample of 88 correctly. In a harder task, where expert human performance is expected to be significantly less than perfect, our system achieves a correct classification rate of over 80%, with clear indications that performance can be improved if a larger dataset were available. Furthermore, very little tuning or adaptation has been necessary to achieve these results giving cause for optimism in the general applicability of this system to other texture classification problems in forensic analysis.

Download activity - last month

Export as XMLCSVJSON

Download activity - last 12 months

Export as JSONXMLCSV

Downloads by country - last 12 months

Export as JSONCSVXML

Archive Staff Only

View Item