maryammeda/apiarist-queen-classifier

Apiarist Queen-vs-Worker Bee Classifier

Binary image classifier (EfficientNet-B0, ~5M params) trained to distinguish queen bees from worker bees on cropped bee images.

Built as part of Apiarist, an offline AI hive inspector for backyard beekeepers, made for the Build Small Hackathon.

Why a dedicated classifier?

Multi-class YOLO detectors fight two problems at once (localize + classify) and queens lose because they're rare and visually subtle. A focused binary classifier on cropped bee images is the right architecture: small, fast, trained specifically for one decision.

Training

• Backbone: efficientnet_b0 (ImageNet pretrained)
• Training data: bee crops extracted from labelled bounding boxes in two Roboflow datasets (Matt Nudi honey bees + Hendricks Ricky bee-project)
• 1,146 queen crops + 29,825 worker crops, balanced via weighted sampling
• Heavy augmentation: rotations, flips, color jitter
• 90/10 train/val split, weighted random sampling for class balance
• AdamW + cosine schedule, mixed precision on a single T4 GPU
• Trained on Modal

Validation metrics

• Accuracy: 0.997
• Precision (queen): 0.991
• Recall (queen): 0.934
• F1: 0.962

Recommended use

Pair with a bee detector (e.g. YOLOv8). Run the detector first, then classify each cropped bee through this model. Threshold queen probability at 0.85 for high-precision flagging.

import torch, timm
from torchvision import transforms

ckpt = torch.load("queen_classifier.pt", map_location="cpu")
model = timm.create_model(ckpt["arch"], pretrained=False, num_classes=2)
model.load_state_dict(ckpt["state_dict"])
model.eval()

tf = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.485,0.456,0.406], [0.229,0.224,0.225]),
])

with torch.no_grad():
    probs = torch.softmax(model(tf(crop).unsqueeze(0)), dim=1)
queen_idx = ckpt["class_to_idx"]["queen"]
queen_prob = probs[0, queen_idx].item()

Caveats

The training distribution leans toward close-up macro photos of bees on honeycomb. Generalization to wide-angle inspection photos (with hands / background visible) is weaker, since YOLO's bee bounding boxes on those photos are often smaller and less precise than the training crops.

License

Apache 2.0. Trained on data released under CC BY 4.0.