Object detection model trained using YOLO v5x, a SOTA object detection algorithm.
The model was pre-trained on the Cashew Disease Identification with AI (CADI-AI) train set (3788 images) at a resolution of 640x640 pixels. The CADI-AI dataset is available via Kaggle and HuggingFace.
You can use the raw model for object detection on cashew images.
The model was initially developed to inform users whether cashew trees suffer from:
- pest infection, i.e. damage to crops by insects or pests
- disease, i.e. attacks on crops by microorganisms
- abiotic stress caused by non-living factors (e.g. environmental factors like weather or soil conditions or the lack of mineral nutrients to the crop).
KaraAgro AI developed the model for the initiatives Market-Oriented Value Chains for Jobs & Growth in the ECOWAS Region (MOVE) and FAIR Forward - Artificial Intelligence for All. Both initiatives are implemented by the Deutsche Gesellschaft für Internationale Zusammenarbeit (GIZ) on behalf of the German Federal Ministry for Economic Cooperation and Development (BMZ).
- Load model and perform prediction:
pip install -U ultralytics
import torch # load model model = torch.hub.load('ultralytics/yolov5', 'custom', path='CADI-AI/yolov5_0.65map_exp7_best.pt', force_reload=True) # Images img = ['/path/to/your/image.jpg']# batch of images # set model parameters # set Non-Maximum-Suppression(NMS) threshold to define # minimum confidence score that a bounding box must have in order to be kept. model.conf = 0.20 # NMS confidence threshold # perform inference results = model(img, size=640) # Results results.print() results.xyxy # img1 predictions (tensor) results.pandas().xyxy # img1 predictions (pandas) # parse results predictions = results.pred boxes = predictions[:, :4] # x1, y1, x2, y2 scores = predictions[:, 4] categories = predictions[:, 5] # show detection bounding boxes on image results.show() # save results into "results/" folder results.save(save_dir='results/')
- Finetune the model on your custom dataset:
yolov5 train --data data.yaml --img 640 --batch 16 --weights KaraAgroAI/CADI-AI --epochs 10
The model has a few limitations that affect its performance in distinguishing between the disease class and the abiotic class. The primary challenge lies in the similarity between these two classes within a typical farm setting. The model may encounter difficulties in accurately differentiating between them due to their overlapping characteristics. This limitation is an inherent challenge in the dataset and can impact the model's accuracy when classifying these classes.
However, it is worth noting that the model exhibits strong performance when it comes to the insect class. This is attributed to the distinct characteristics of insect class, which make them easier to identify and classify accurately.
If you want to know how the model and dataset has been used further for the GIZ-funded activity, please have a look at:
- The GitHub repository for the CADI AI desktop application
- Downloads last month