Fig 1.
General workflow diagram.
Fig 2.
Sample images for each class ((a) Anthracnose, (b) Bacterial Canker, (c) Cutting Weevil, (d) Die Back, (e) Gall Midge, (f) Powdery Mildew, (g) Sooty Mould, (h) Healthy).
Table 1.
Hardware specifications for training and inferencing.
Table 2.
Comparative model performance metrics (Cross validation 5 fold).
Fig 3.
Deep learning model cross-validation result.
Table 3.
10-Fold cross-validation performance summary for CNN models (GPU).
Fig 4.
Deep learning model cross-validation result.
Table 4.
Paired t-test results between CNN models.
Fig 5.
Confusion matrix MobileNet (initial).
Fig 6.
Grad-CAM illustrating the model’s focus for healthy and diseased leaves.
Table 5.
Comparative analysis of features for healthy and diseased leaves (based on Grad-CAM insights).
Fig 7.
Anthracnose Gradcam.
Fig 8.
Bacterial Canker Gradcam.
Fig 9.
Cutting Weevil Gradcam.
Fig 10.
Die back Gradcam.
Fig 11.
Gall Midge Gradcam.
Fig 12.
Powdery Mildew Gradcam.
Fig 13.
Sooty Mould Gradcam.
Fig 14.
Machine learning algorithms performance (cross validation).
Table 6.
10-Fold cross-validation and test performance for classical ML models.
Fig 15.
Confusion matrices for the RandomForest, SVC, and logistic regression models.
Table 7.
Paired t-test results between ML models.
Fig 16.
Hybrid model workflow diagram.
Table 8.
Comparative analysis of preprocessing time and iterations.
Fig 17.
Model performance (controlled use-case).
Table 9.
5-Fold cross-validation results on test dataset (controlled use case GPU).
Table 10.
Paired t-test results comparing MobileNet and hybrid model.
Table 11.
Performance comparison for random use case across devices.
Table 12.
Per-image inference time and time reduction for hybrid model vs MobileNet.
Fig 18.
Confusion matrix Mobilenet model (test data).
Fig 19.
Confusion matrix Mobilenet model (Test data).
Table 13.
Performance of different devices at varying levels.
Table 14.
Performance comparison between hybrid model and MobileNet (test set).
Table 15.
Performance comparison between hybrid model and MobileNet (without Rembg and CLAHE).
Fig 20.
Confusion matrix Mobilenet model (test data).
Fig 21.
Confusion matrix hybrid model (test data).
Fig 22.
Prototype interface of the application used for real-time disease detection in mango leaves.