TY - GEN
T1 - Statistical modeling, detection and segmentation of stains in digitized fabric images
AU - Gururajan, Arunkumar
AU - Sari-Sarraf, Hamed
AU - Hequet, Eric F.
PY - 2007
Y1 - 2007
N2 - This paper will describe a novel and automated system based on a computer vision approach, for objective evaluation of stain release on cotton fabrics. Digitized color images of the stained fabrics are obtained, and the pixel values in the color and intensity planes of these images are probabilistically modeled as a Gaussian Mixture Model (GMM). Stain detection is posed as a decision theoretic problem, where the null hypothesis corresponds to absence of a stain. The null hypothesis and the alternate hypothesis mathematically translate into a first order GMM and a second order GMM respectively. The parameters of the GMM are estimated using a modified Expectation-Maximization (EM) algorithm. Minimum Description Length (MDL) is then used as the test statistic to decide the verity of the null hypothesis. The stain is then segmented by a decision rule based on the probability map generated by the EM algorithm. The proposed approach was tested on a dataset of 48 fabric images soiled with stains of ketchup, corn oil, mustard, ragu sauce, revlon makeup and grape juice. The decision theoretic part of the algorithm produced a correct detection rate (true positive) of 93 % and a false alarm rate of 5% on these set of images.
AB - This paper will describe a novel and automated system based on a computer vision approach, for objective evaluation of stain release on cotton fabrics. Digitized color images of the stained fabrics are obtained, and the pixel values in the color and intensity planes of these images are probabilistically modeled as a Gaussian Mixture Model (GMM). Stain detection is posed as a decision theoretic problem, where the null hypothesis corresponds to absence of a stain. The null hypothesis and the alternate hypothesis mathematically translate into a first order GMM and a second order GMM respectively. The parameters of the GMM are estimated using a modified Expectation-Maximization (EM) algorithm. Minimum Description Length (MDL) is then used as the test statistic to decide the verity of the null hypothesis. The stain is then segmented by a decision rule based on the probability map generated by the EM algorithm. The proposed approach was tested on a dataset of 48 fabric images soiled with stains of ketchup, corn oil, mustard, ragu sauce, revlon makeup and grape juice. The decision theoretic part of the algorithm produced a correct detection rate (true positive) of 93 % and a false alarm rate of 5% on these set of images.
KW - Expectation-maximization
KW - Gaussian mixture model
KW - Minimum description length
UR - http://www.scopus.com/inward/record.url?scp=34548247822&partnerID=8YFLogxK
U2 - 10.1117/12.705105
DO - 10.1117/12.705105
M3 - Conference contribution
AN - SCOPUS:34548247822
SN - 0819466166
SN - 9780819466167
T3 - Proceedings of SPIE - The International Society for Optical Engineering
BT - Proceedings of SPIE-IS and T Electronic Imaging - Machine Vision Applications in Industrial Inspection XV
PB - SPIE
T2 - Machine Vision Applications in Industrial Inspection XV
Y2 - 29 January 2007 through 30 January 2007
ER -