High-Precision Real-Time Wheat Head Counting via Enhanced Feature Pyramids

Accurate and real-time detection of wheat heads is a critical task for yield estimation and precision agriculture. A novel wheat head detection framework is proposed in this paper. This framework is based on Detectron2 architecture and integrates a multi-stage preprocessing pipeline that includes CLAHE (Contrast Limited Adaptive Histogram Equalization)-based contrast enhancement, Gaussian filtering, and median filtering to improve input quality under real-world conditions. The model leverages a Residual Network and Feature Pyramid Network (ResNet + FPN) backbone to extract multi-scale features and employs a Region Proposal Network (RPN) with Region of Interest (ROI) align to refine bounding box localization and object classification. The proposed model achieves a mean Average Precision (mAP) of 90.25%, a precision of 88.7%, recall of 85.1%, and an F1-score of 0.874, outperforming state-of-the-art detectors such as You only live once YOLO version 5 (YOLO v5), YOLOv7, EfficientDet-D3, Cascade Region-based Convolutional Neural Network (R-CNN), and Swin Transformer. The system achieves real-time inference at 9.75 frames per second (FPS) with a model size of 176 MB and inference time of 95 ms per frame. Statistical analysis across five independent runs shows a low standard deviation (±0.18) and a significant p-value (<0.01) confirming result consistency. The proposed architecture offers high detection performance and stability, demonstrating its suitability for deployment in smart farming applications.