Abstract:
Ethiopia is the birthplace of coffee, where it is naturally gifted and a major source of
foreign income. In Ethiopia, coffees are cultivated in different regions but the dominant
regions are Harar, Jimma, Limu, Sidama, and Wellega. The different varieties of coffees
are circulated domestically and export to international market based on their specific
origin manually. However, maintaining and assuring the quality of coffee based on their
specific origin manually is a worthful task; because it is highly biased, subjective and
prone to error. As a result, automation of these manual coffee bean varieties identification
and classification based on their growing region is required. The main objective of this
study is designing and developing an automatic Ethiopian coffee bean varieties
identification based on their growing regions using image processing and machinelearning
techniques. To achieve the objective of this study, images of coffee beans are
collected by taking different coffee samples of Harar, Jimma, Limu, Sidama and Wellega
from Ethiopian Commodity exchange (ECX), Addis Ababa. Following the collection of the
dataset, the developed system incorporates various processing phases such as image
resizing, image filtering, image contrast enhancement, noise removal, grayscale
conversion for preprocessing and a combined approach of thresholding and K-means
segmentation technique for grayscale and RGB image respectively. After, preprocessing
and segmentation of images, we have used an end-to-end CNN model using SoftMax and
RBF kernel function of SVM for classification of Ethiopian coffee bean varieties. We
compared the proposed end-to-end CNN to VGG16 and ResNet50, and the proposed endto-
end CNN had the highest accuracy of 89.99% than VGG16 (87.75%) and ResNet50
(81.25%). Even if, the proposed end-to-end CNN model achieved 89.99% accuracy but the
performance of the model is affected by similarity of color features and we investigated to
merge the color feature with texture feature by applying local feature descriptor using
HOG. We evaluated the performance of HOG feature extraction, CNN feature extraction,
and a hybrid feature vector (HOG-CNN) by using a multi-class SVM classifier, in which
74.17%, 85.83% and 97.5% accuracy achieved respectively. Finally, the deep-shallow
based feature (CNN-HOG) are registered with highest accuracy of 97.5 % in this study.
Keywords: CNN, HOG, HOG-CNN, PCA, SVM, SoftMax, K-means, Thresholding
