TWO-STAGE HYBRID YOLO-EFFICIENTNET FRAMEWORK FOR APPLE DISEASE DETECTION
DOI:
https://doi.org/10.5281/zenodo.24271939Abstract
Accurate recognition of apple fruit diseases from orchard images is difficult because fruits may be partly occluded and disease symptoms can be visually similar. This paper presents a two-stage hybrid vision pipeline that separates fruit localization from disease classification. The first stage uses a fine-tuned YOLO11n detector, optimized through a 50-iteration genetic hyperparameter search, to localize apple regions in fullscene images. Detected regions are cropped, resized to 224 x 224 pixels and passed to a pre-trained EfficientNetB0 classifier. To reduce class-frequency bias, the classifier backbone is unfrozen and fine-tuned for ten epochs using class-weighted cross-entropy. The dataset combines 4,154 images from two Roboflow repositories and represents eight classes: Anthracnose, Black Pox, Black Rot, Blotch, Healthy, Powdery Mildew, Rot and Scab. Fine-tuned YOLO11n achieved 94.22% precision, 92.60% recall, 96.86% [email protected] and 78.63% [email protected]:0.95, with detector latency of 3.0 ms per image on an NVIDIA GeForce RTX 2060. Fine-tuned EfficientNet-B0 reached 96.49% peak validation accuracy. The findings indicate that separating localization and classification is a promising approach; independent field testing and end-to-end latency measurement are still required.
Downloads
Published
Issue
Section
License

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.













