Authors

Abstract

Real-time, high-precision detection of wheat grains is crucial for food security and intelligent management, yet fully supervised methods require extensive annotations and struggle with occlusion and overlap. This paper proposes a lightweight YOLOv8-CoT model based on EfficientTeacher. FasterNet is integrated with CoTAttention to optimize the FC-C2f unit, enhancing channel–spatial feature representation, while a CBAM module is inserted at the end of the neck to improve recognition of occluded and overlapping grains. A pseudo-label self-training strategy is adopted using 80% unlabeled data and 20% labeled samples. The proposed method achieves 91.7% accuracy in field scenarios, improves efficiency by 6.6%, and reduces annotation cost to one-fifth.

Abstract in Chinese

小麦籽粒实时高精度检测对粮食安全与智能管理至关重要，但全监督方法依赖大量标注且难应对遮挡重叠。本文提出基于 Efficient Teacher 的轻量化 YOLOv8-CoT：融合 FasterNet 与 CoTAttention 优化 FC-C2f，提升通道-空间特征表征；在颈部引入 CBAM 增强遮挡与重叠识别；采用 80% 未标注与 20% 标注数据进行伪标签自训练。在田间场景实现 91.7% 精度，效率提升 6.6%，标注成本降至 1/5。