| Scope and method of study. Image processing techniques were developed for grading dates into quality classes on the basis of color and texture. Three approaches: first-order histogram (FOH), co-occurrence matrix (CCM), and texture spectrum (TS) were applied to dates which had been manually classified according to USDA and Industry grading standards. Color images in HSI and RGB were acquired from each date. Regions-of-interest were processed from four locations on the date surface. A total of 132 features were extracted from each region-of-interest (ROI), twenty-two from each of six color bands (H, S, I, R, G, and B). Eighteen models were developed. Each model was composed of subsets of features. Two models used features of the FOH, eight used features of the CCM, and eight used features of the TS. Feature observations for each model were classified into grades using multivariate discriminant procedures.; Findings and conclusions. Classification accuracy varied among models. Highest accuracy was 65.8%, 98.4%, and 84.4% from the FOH, CCM, and TS, respectively, using USDA grading standards. Accuracy increased to 77.6%, 99.3%, and 89.0% with Industry grading standards. There was no significant difference in classification accuracy among observations obtained from regions-of-interest at different locations on the date. There was also no significant difference between RGB and HSI models. Accuracy of RGB models extracted from the CCM and TS was slightly higher than that of the HSI models. RGB models obtained from the FOH were less accurate than the HSI models. Highest accuracy of black-and-white models represented by the intensity band of the HSI color system was 77%. Accuracy of the CCM models was significantly greater than that of the TS models. |
