<![CDATA[Early detection of anthracnose disease on papaya leaves is important for mitigating crop yield losses, but manual methods are inefficient and prone to subjectivity; this study evaluates the effect of region of interest (ROI) extraction strategies on deep learning-based classification performance. The objective of this study is to compare four classification pipelines: ResNet-50 without segmentation, (M1) ExG+Otsu + ResNet-50, (M2) U-Net + ResNet-50, and (M3) RCNN + ResNet-50, in detecting anthracnose on papaya leaf images. The methods included the use of the public BDPapayaLeaf dataset, pre-processing and augmentation, and evaluation using stratified K-fold cross-validation with evaluation metrics including precision, recall, and F1. The results show that the semantic segmentation-based pipeline, U-Net + ResNet-50 (M2), provides the best performance with an F1-score (macro/weighted) ≈ 0.961 and precision–recall balance in both classes; M0 showed the highest recall for the anthracnose class (≈0.99) while M1 based on color thresholding provided the lowest performance due to sensitivity to lighting variations; M3 (RCNN) was in the middle. The findings recommend the use of segmentation and classification pipelines for field applications, noting the need for dataset expansion and inference optimization for deployment.]]>
Copyrights © 2026
