Abstract:
Eragrostis Teff is a tiny, round grain and it is the smallest cereal grain with an average length of 1.0 mm and a width of about 0.60 mm. It replaces gluten-containing flours such as regular wheat flour. It is one of the most widely used grains for baking injera and it is the most commonly used grain for nutrition specifically in Ethiopia and the neighboring country Eritrea. Compared to the other region, Teff cereal is produced predominantly in Amhara and Oromia regions. More specifically, in the Amhara region, Adiet, Bichena, Debre Markos, and Dejen are well-known Teff grain-producing areas. Due to the tiny nature of Teff cereal and closer environmental behavior, it is difficult to make certain and ensure the intended cereal product quality and source using traditional mechanisms. The proposed study thus focuses on developing production area and quality classification of Teff cereal using machine learning approach. The data was collected from Adiet, Bichena, Debre Markos, and Dejen, areas consisting based on their production areas while two types of cereal (high and low quality) taking the respective quality in each growing area, collectively eight classes were considered. For this study, 740 images were taken for each class and collectively a total of 5,920 images were collected. After data collection, image enhancement techniques such as image scaling, data augmentation, histogram equalization, and noise removal are used. After that, CNN was used to extract interesting features by lowering the dimensionality of the data. From the total of images, 80% of images were used for training, while 20% of the images are used for testing the model. Finally, FVGG16, FINCV3, QSCTC, and EMQSCTC with Softmax classifier are used to classify the image into the pre-defined class. Besides, SVM and RF machine learning classifiers are also used by taking the features extracted using CNN. As a result, the accuracy of FVGG16, FINCV3, QSCTC, EMQSCTC, SVM and RF are 93.92%, 93.24%, 95.27%, 97.72%, 90.29% and 86.91 % respectively. Accordingly, the ensemble of FVGG16, FINCV3, and QSCTC using the Max-Voting approach performs better than individual algorithms. This study needs furthermore improvement by considering other types of Teff cereal (Qey and Mixed Teff) and following the way of hierarchical classification.
Keywords: Teff, Ensemble Learning, Max-Voting, CNN, SVM, and RF
