Abstract:
This research aims to compare the efficiency of techniques for classifying rice varieties from images of milled rice grains. Five rice varieties were considered: Karacadag, Jasmine, Ipsala, Basmati, and Arborio. Image processing combined with machine learning methods were applied. The procedure started with image processing to reduce noise from the images of rice grains of various varieties, which were color JPEG format images with a resolution of 250x250 pixels, with a total of 15,000 images per variety. All noise-reduced images were then processed for classification using seven different techniques: Canny edge detection, Sobel edge detection, ridge detection, texture detection, image enhancement with Laplacian filters, image enhancement with Gaussian blur, and histogram equalization. Features, including 21 shape features and 11 texture features, were extracted and classified using six machine learning techniques: decision trees, Naïve Bayes, k-Nearest Neighbors, Artificial Neural Network (ANN), Support Vector Machines (SVMs), and gradient boosted trees. Training was conducted with K-fold cross-validation with K=10 for all machine learning techniques. The research findings showed that using image processing with Sobel edge detection combined with classification using SVMs was the most effective method, with classification accuracies of 98.68%, precision of 98.67%, recall of 98.67%, F1-score of 98.67%, and a Cohens kappa coefficient of 98.35%. The classification process took 136.21 seconds.