f4m1/plant-disease-detector-12

Plant Disease Detector (12 classes)

Frozen YOLO11s detector used by the plant diagnosis API. It detects disease types independently of crop identity; the companion classifier supplies the crop prediction. The API uses the pinned best.pt checkpoint.

Classes

powdery_mildew, early_blight, late_blight, leaf_spot, bacterial_leaf_spot, phytophthora_blight, leaf_rust, stem_rust, stripe_rust, septoria_leaf_blotch, fusarium_head_blight, downy_mildew.

Inference

from ultralytics import YOLO

model = YOLO("best.pt")
results = model.predict("leaf.jpg", imgsz=640, conf=0.20, iou=0.70)

The API defaults to confidence 0.20, IoU 0.70, image size 640, and at most 300 detections. A validated dynamic-shape ONNX export is also available as best.onnx, but it is not used by the API. On the held-out grouped test split, standard NMS at the training evaluation settings produced precision 0.6743, recall 0.6050, mAP50 0.6277, and mAP50-95 0.4613. A separate inference sweep found confidence 0.20 / IoU 0.70 gave recall 0.6568 and mAP50 0.6329.

Limitations

This is a research model trained from heterogeneous public and field imagery. Predictions are not a professional diagnosis. Performance varies substantially by disease class and image domain. The training data licenses and provenance remain recorded in the source project and are not relicensed by this model repository.