Abstract:
Crop disease is a significant issue for Bangladesh's economy, but it can be prevented with early detection. This thesis proposes a deep learning-based instance segmentation technique for detecting 10 of the most common crop diseases in Bangladesh. This technique can enable automated crop disease detection on a large scale. The paper introduces a new annotated dataset of around 4600 images for 10 different disease classes. Image annotation is usually the most time- consuming phase for any segmentation task. To reduce this time, the paper proposes a new semi- automated annotation pipeline. It showcases how this pipeline can reduce image annotation time by approximately 85%. After annotating all the images, three different models were trained – two variants of YOLOv8 (YOLOv8, YOLOv8l) and YOLOv9c. These models were trained on three different versions of the dataset: the first version with only manually annotated images, the second version with a combination of manually and semi-automated annotated images, and the third version with an augmented combined dataset. The augmented version did not perform well, but when increasing the dataset using semi-automated annotation, the mAP score increased. YOLOv8l achieved the best mAP score of 0.7417.
