Intelligent rice quality assessment using hybrid CNN-clustering approach

While rice quality assessment can be performed through various approaches, this paper focuses specifically on broken rice percentage detection, which is critical for trade compliance and international export standards. Current methods face two primary limitations: reliance on fixed thresholds to classify broken versus head rice, which fails across different rice varieties, and fragment counting approaches that inflate broken rice percentages by multiple counting fragments originating from the same rice kernel. This paper presents a novel integrated approach combining unsupervised K-means clustering with supervised deep learning for enhanced rice grading accuracy. The system uses K-means clustering with morphological features to automatically separate head rice from fragments without arbitrary thresholds, addressing variability across rice varieties. A convolutional neural network (CNN) then classifies identified fragments into three morphological categories—Hooked Top (HT), Rectangular Midsection (RM), and Rounded Bottom (RB)—achieving 98.95% accuracy. The key innovation lies in a reconstruction algorithm that intelligently recombines complementary fragment pieces to determine true broken rice rate (BRR), unlike conventional methods that overestimate by counting fragments individually. Size proportions (1/3, 1/2, 2/3) are adaptively determined relative to each sample's head rice dimensions from clustering. The CNN architecture features three convolutional blocks with optimized regularization, trained on 300 original basmati rice images systematically augmented to 5,700 samples through rotation, scaling, and flipping techniques, with 75% allocated for training and 25% for testing. Experimental results demonstrate consistent performance and superior accuracy in automated quality assessment for precise rice grading.