Abstract:
Agriculture plays a key role in the food security and economic development of developing countries,
including Ethiopia. However, accurately identifying the geographical origin of agricultural products,
such as Teff, can be challenging due to similarities in size, color, and shape. While previous studies have
explored classification methods, the physical properties of Teff pose segmentation, noise, and
classification challenges for traditional techniques. This study uses a multi-channel capsule network and
computer vision to identify the geographical origin of Ethiopian Teff, including red and mixed varieties
from six regions. The approach addresses issues such as viewpoint variance, overfitting, and spatial
invariance faced by previous convolutional neural network methods. Teff images were collected from
various regions in the Amhara region of Ethiopia, including East Gojjam, West Gojjam, North Gonder,
and North Shoa, with a specific focus on locations such as Adet, Bichena, Dejen, Denbya, Debre Markos,
and Minjar. A total of 2700 Teff images were acquired using a smartphone camera. To enhance the
image quality, we used various techniques, including denoising filters such as median, Gaussian, and
non-local means. We also applied Content-aware resizing to adjust the image dimensions accordingly.
Additionally, Contrast Limited Adaptive Histogram Equalization (CLAHE) was used to enhance the
images further. The next step involved image segmentation, where we used a combination of multi-Otsu
thresholding, region-based segmentation, watershed techniques, and the U-net method, which is widely
used in deep learning applications. Subsequently, we extracted features from the segmented images
using the BRISK and HOG methods. Once the feature vectors were obtained, we developed a model using
Capsule Networks to determining the geographical origin of Ethiopian Teff. To train and evaluate the
model, we divided the dataset into training and testing sets, allocating 90% of the data for training and
10% for testing the model's performance. By conducting 11 experiments, we achieved notable results in
three different approaches. The first approach, using the BRISK feature extraction method, combined
Multi-Otsu thresholding, region-based segmentation, watershed segmentation, and comprehensive noise
removal, leading to the highest accuracy of 84%. In the second approach, combining region-based
segmentation with median and Gaussian filtering, along with BRISK feature extraction, resulted in an
accuracy of 80%. The third approach, using Histogram of Oriented Gradients (HOG) feature extraction,
combined Multi-Otsu thresholding, region-based segmentation, watershed segmentation, and
comprehensive noise removal, achieving an impressive accuracy of 92.5%. These findings demonstrate
the successful development of a highly accurate model for determining the geographical origin of
Ethiopian Teff, with a test accuracy rate of 92.5%.
Keywords: Multi-Channel, Capsule Network, Computer vision, Geographical Origin, Teff,
Deep Learning, Multi-Otsu, HOG, BRISK