<title>Nonlint material identification using computer vision and pattern recognition</title>

Abstract

ABSTRACT This paper discusses methods used to evaluate a feature space for identification of non-lint mate-rial (trash) in cotton samples. A main criterion for accepting any feature in the identification task was invariance under translation, rotation, and, in most cases, scale. In subsequent processing, most features were normalized. Classical grouping was performed in an n-dimensional featurespace using divisive hierarchical clustering based on the Euclidian distance metric. The best re-sults for identifying bark, stick, and leaf/pepper trash in the sample data set was 92%. By cate-gory, bark was identified correctly 88%, stick 84%, and leaf/pepper 94% of the time. Identifica- tion between leaf and pepper could be handled by defining an area cutoff in the pepper-leaf continuum. 2. BACKGROUND The Southwestern Cotton Ginning Research Laboratory (SWCGRL) is part of the United StatesDepartment of Agriculture (USDA), Agricultural Research Service. SWCGRL is charged withinvestigating any aspects of cotton ginning where improvement could result in better return for the