Description of Invention
Classifying and detecting plant diseases from leaf images captured in real-world environments present significant challenges, primarily due to the overlapping of leaves and the presence of complex backgrounds. To mitigate these issues, up to five target leaves per image were annotated in the Cassava 2020 dataset, published by Makerere AI Lab. These annotations were utilized to train a Faster RCNN model with ResNet101 for the automatic detection of diseased leaves. The detected leaves were then isolated from the background through a masking process, resulting in masked images Mi, where i is from 1 to 5, indicates the number of isolated leaves in the masked image. These masked images were subsequently used to train various CNN models—EfficientNetB1, DenseNet121, ResNet50, and Xception—for cassava disease classification. The results demonstrated that models trained on M? images achieved an accuracy improvement of 2.13% to 3.06% compared to those trained on the original images.