RiceNet: A deep convolutional neural network approach for classification of rice varieties

Abstract

The cultivation of desired grain varieties holds immense significance as about 67% of the world’s population is associated with the agriculture sector. Unknowingly sowing the wrong variety of seeds may lead to a colossal waste of effort and money. Furthermore, the growing issue of rice grain adulteration in high-quality rice poses a threat to the trust of rice importers and exporters. While traditional methods are expensive, laborious, and error-prone, Computer Vision provides a good alternative that constitutes a current, advanced technology for image processing and data evaluation that holds tremendous promise and potential. In this research study, five varieties of rice grains including Jehlum Sr-1, Mushkibudji, Sr-2, and Sr-4 were collected from local grain and were used for research analysis. A computer vision system “RiceNet” contingent upon Deep Convolutional Neural Network (DCNN) framework has been designed for ameliorating the accuracy of identifying five unique groups of rice grain varieties. Deep Learning (DL) based pre-trained architectures including InceptionV3 and InceptionResNetV2 models were also adopted for classifying five specific groups of rice species. To optimize model parameters and alleviate back-propagation error during training, the Adam optimizer with a learning rate (lr) of 0.00003 has been employed to fine-tune the pre-trained InceptionV3 and ResNetInceptionV2 models. The proposed RiceNet architecture and pre-trained models were also compared with traditional ML approaches of HOG-SVM, SIFT-SVM, HOG-Logistic Regression(HOG-LR), SIFT-Logistic Regression(SIFT-LR), HOG-KNN, and SIFT-KNN for rice grain classification. With these experimentations at hand, it was observed that our proposed model “RiceNet” outperformed other approaches in similar computer vision tasks. The prediction accuracy outcome for the test dataset by HOG-SVM, SIFT-SVM, HOG-LR, SIFT-LR, HOG-KNN, and SIFT-KNN models were 66.0%, 65.33%, 62.67%, 65.0%, 54.0%, and 52.0% respectively. RiceNet, InceptionV3 and ResNetInceptionV2 have the best prediction accuracy of 94%, 84% and 81.333%. The remarkably high success rate of DCNN models makes them highly valuable and can be extended to endorse an integrated grain identification system that can operate in real-world situations.