Cotton is currently graded on its color, leaf, and preparation. Individual component measurements for color and trash are currently reported along with the cotton grades. The cotton industry now requires a method to identify all types of trash in a sample. An image processing system, a test of feature invariance, and two preliminary pattern classification methods used to identify the types of trash in cotton samples are presented. In one approach, the classical grouping performed uses divisive hierarchical clustering based on a normalizecj Euclidian distance metric. Clustering classified 568 trash objects in the training data set with 92% accuracy into bark, stick, and leaf/pepper categories. An area cutoff in the pepper-leaf continuum could handle a separation between leaf and pepper. In the second approach, neural networks minimize identification error in a training data set. Using 75% of a 562 object data set for training, neural networks classified the remaining 134 trash objects into bark versus nonbark categories with a 99.3% accuracy (one piece of bark misidentified). Another network further classifies the nonbark objects into either stick or a combined leaf/pepper category with 96.4% accuracy. Combining both networks gives an accuracy of 96.3%.