It is well known that disagreements in cotton color grades between the high volume instrument and classer are substantial. The machine-classer disagreement deters the full acceptance for the use of machine grading of cotton color. This paper first provides a quantitative analysis on the distributions of the disagreements across all the color grades, the major and sub-color categories. The study proves that the disagreements can be both systematic and random, and further analyzes the possible sources for these two types of disagreements. The paper devotes its second part to the introduction of a novel design of a neural network classifier for cotton color classification. This classifier consists of multiple networks performing a two-step classification that identifies the major and sub-color categories separately. The classifier can be trained by any desirable data. In this research, it was trained by using a set of classers' grades, and exhibited good generalization for the new testing data. The classifier seems to have reduced the machine-classer disagreements to a minimal level, which is limited by the classer's sustainability.

According to the USDA universal standards for Upland cotton, cotton colors are classified into five major categories based on chromatic differences (1-white, 2-light spotted, 3-spotted, 4-tinged and 5-yellow stained), and three to eight subcategories in one major category based on differences in grayness (1-good middling, 2-strict middling, …8-below grade) [3]. A double-digit number that indicates both the major and sub-categories of the color is used to denote a color grade. For example, color grade 21 refers to a white, strict-middling cotton. The color grade of a sample is determined either by a classer who compares the sample with the universal standards; or by the colorimeter of a high volume instrument (HVI) that calculates the location of the color data of the sample in the cotton color diagram. Although the HVI is a unique instrument currently used for grading cotton colors in the cotton classing system, its output has not been accepted as official color grading by the industry because of substantial disagreement with grades provided by a classer. A classer, who is trained to visually grade cotton color and trash, has the right to correct the HVI's rating when a dispute occurs. Since visual grading has been the traditional and widely accepted method for cotton color grading, the machine-classer disagreement undermines the industrial acceptance for the machine grading. To investigate new methods that can reduce this disagreement, it is necessary to understand the possible reasons causing the disagreements. In this paper, we will first report a study on the HVI-classer disagreements in color grades, and then present the preliminary results of using a neural network classifier to reduce these disagreements.