Peonies Under the Lens: How Does AI Learn to Tell Every Variety Apart?

The peonies of Luoyang, China, are world-famous for their beauty and economic value. But telling one variety from another out in the fields is tricky, even for experts. The flowers look very similar, often overlap, and outdoor photos are rarely perfect. That's why a research team turned to artificial intelligence. They started with YOLOv5, a well-known object detection model, and gave it a serious upgrade. The result is called BCP-YOLOv5, a name that packs in three key innovations: Biformer, an attention system that helps the algorithm stay focused even when leaves get in the way; CARAFE, which improves image quality during analysis; and Focal-CIoU, a function that learns well even from imperfect photos. In field tests, the new model improved precision by 3.4%, recall by 4.4%, and overall accuracy by 4.5% compared to the original version. A small step for algorithms, a big help for farmers who want to monitor their crops with smart tools.

The BCP-YOLOv5 model was designed to solve a very specific problem: automatically recognizing peony varieties in real-world environments, where lighting conditions, occlusions, and similarity between cultivars make analysis difficult. The architecture is based on YOLOv5s, one of the lightest versions of the YOLO (You Only Look Once) family, known for its real-time object detection capabilities. The researchers introduced three key modifications. The first is Biformer, an attention module based on Vision Transformers. Unlike traditional attention mechanisms that compute relationships between all pixels, Biformer uses a bi-level approach: it first selects the most relevant areas of the image, then computes attention only on those. This drastically reduces computational load and improves the model's ability to focus on partially hidden flowers. The second modification involves CARAFE (Content-Aware Reassembly of Features), an upsampling operator that replaces classical methods like nearest neighbor. CARAFE analyzes the local content of the image and reconstructs details more faithfully, which is essential when working with small or partially covered flowers. Finally, the loss function was redesigned drawing inspiration from Focal-EIoU: the new Focal-CIoU gives greater weight to well-localized samples, reducing the influence of noisy or imprecise bounding boxes during training. The experimental results, conducted on a Luoyang peony dataset, show a significant improvement over YOLOv5s, with mAP@0.5 reaching 93.2%. This approach paves the way for automatic monitoring systems in precision agriculture, where recognizing cultivated varieties is essential for traceability and quality management.