Abstract:

In order to assess preharvest cotton grades, according to Chinese government grading standard, clustering fusion was performed based on machine vision technologies by using K-means and competitive learning network to grade cotton quality with 7 categories renewedly based on their size, white, impurity and yellow in Ohta, HIS, and Hunter Color Spaces. Correlation analysis showed that the Peason’s correlations among image features were significant at the 0.01 probability level by adjusting image intensity and Hunter Color Space to an approximate optimum; clustering by human eyes did not consider all image features uniformly with fitting coefficients of quadratic polynomial of 0.55–0.98 (0.88, 0.94, 0.98, 0.55) between cluster center of image features and grade value of cotton quality; individual clustering by K-means and competitive learning network also did not consider all image features uniformly with fitting coefficients of 0.32–0.74 (0.74, 0.63, 0.70, 0.32) and 0.39–0.94 (0.85, 0.39, 0.94, 0.84), respectively; and their clustering fusion considered all image features uniformly with fitting coefficients of 0.71–0.99(0.89, 0.71, 0.99, 0.83). Bayes quadratic discriminants analysis for cotton graded showed that clustering by human eyes recognized the 1st, 2nd, 7th grades with accuracies of 73%–100%, the grades 3–6 with accuracies of 26%–46%, and total accuracy of 47.7%; accordingly, clustering by K-means recognized each grade with accuracies of 93%–100%, and total accuracy of 96%; clustering by competitive learning network recognized each grade with accuracies of 79%–95%, and total accuracy of 86%; clustering fusion recognized each grade with accuracies of 65%–100%, and total accuracy of 78.6%. On the whole, the average quality grade of clusering fusion was 4.33 while that of clustering by human eyes was 4.57, and the specimens with large recognization difference between the two methods were less than 1/3 of the total. Compared with by human eyes, clustering fusion can use each image feature more adequately and uniformly with the wider range of vision based on human’s previous knowledge, and overcome the over-training of individual clustering, further, grade preharvest cottons objectively to improve high-quality cottons to be purchased, and this method can be generalized effectively to meet different environments.

Key words: Cotton grade, Machine vision, Image features, Cluster fusion, Accuracy