This study addresses the problem of wheat seed pose detection prior to slicing and proposes an improved YOLO-WP (YOLOv8-WheatPose) model. Built upon the YOLOv8n framework, the method detects seed embryos and awns, calculates the embryo orientation vector from bounding box centers, and determines pose angles using the arctangent function. To enhance performance, we integrate the C2fFaster module for optimized cross-stage connections, employ an Efficient Multi-scale Attention (EMA) mechanism for improved feature representation, and redesign the neck network with a Bidirectional Feature Pyramid Network (BiFPN) for effective multi-scale fusion. Experimental results show that YOLOv8-WP achieves 81.8% AP in embryo and awn detection, reduces parameters by 23.33%, maintains a model complexity of 6.1 GFlops, and achieves an average angular error of less than 2°.

Wheat seed，YOLOv8，C2fFaster module ，Efficient Multi-scale Attention，Bidirectional Feature Pyramid Network