Authors

Abstract

In contemporary agricultural practices, the use of image and video acquisition technologies, such as drones and cameras, has become increasingly common for capturing and monitoring crop growth in agricultural fields. The reliance on visual data for analyzing farm management conditions and facilitating decision-making processes is gaining significant traction. However, in practical applications, image acquisition tools often face challenges in maintaining optimal distance and angle during data capture, which can negatively impact the detection accuracy of existing object detection methods. Semi-supervised learning plays a crucial role in improving object detection. In this study, a semi-supervised algorithm for wheat spike recognition was developed based on an optimized YOLOv8n model. The model incorporates SPDConv and PSA attention modules after the SPPF layer, effectively reducing computational and memory demands while enhancing model performance. The proposed model achieved an accuracy of 94.2%, outperforming YOLOv5s, Efficient Teacher, and the baseline YOLOv8n by 10.9%, 4.5%, and 6.1%, respectively—demonstrating its strong potential for practical agricultural applications.

Abstract in Chinese

当今农业，使用无人机或摄像头这样的图像视频采集工具拍摄或监控农田作物生长情况变得常见，依靠这些图像分析农田管理情况和辅助决策也会变得越来越有市场。但在实际生产中，图像采集工具在拍摄时很难保持合理的距离和角度，这使得现有的目标检测方法检测精度低。而半监督学习对提高物体检测至关重要。我们开发了一种用于麦穗识别的半监督学习算法，使用了改进的YOLOv8n，并在SPPF之后集成了PSA注意力机制和SPDConv。这些改进减少了计算和内存需求，提升了模型性能。我们的模型达到了94.2%的准确率，分别比YOLOv5s、Efficient Teacher和YOLOv8n高出10.9%、4.5%和6.1%，展示了其实际潜力。