Abstract:
There is a wide range of dry seeds in the world. The quality of seeds significantly impacts
crop production. Hence, seed classification is vital for both marketing and production,
serving as the foundation for sustainable principles in agriculture systems. The primary
objective of this study is a method for extracting multiple seeds from a single image by
segmenting them. Thus, a multi-label segmentation and detection system was developed
to distinguish 18 different varieties of seeds with similar features. For the training model,
images of 100 seeds of 18 varieties were taken with a high-resolution camera. In the
initial stage, images undergo augmentation and feature extraction processes. Resulting a
total of 1029 by 18 shapes of labels. The initial phase was focused on processing data and
extracting the valuable features from it. For segmentation and detection, the model
YOLOv8 was applied. Also, to compare the performance of our custom train YOLOv8
model we trained a vanilla YOLOv5. Our breakthrough YOLOv8 model achieved an
impressive 99% confidence level for select classes in our private dataset, with an overall
average detection rate of 94.5%. Whereas the YOLOv5 conducted an overall detection
rate of 94.7%.
