Accurate detection of wheat seedlings is crucial for monitoring early population establishment and evaluating sowing quality. However, detection in real field environments remains challenging due to diverse seedling morphology, varying planting densities, occlusion, and complex background interference. Although deep learning has promoted the development of agricultural vision systems, existing wheat seedling detection methods still suffer from two key limitations: (1) insufficient modeling of spatial contextual relationships, leading to degraded accuracy under dense planting and complex field conditions; and (2) difficulty in balancing detection performance and computational efficiency, restricting real-time deployment on resource-limited agricultural devices. To address these issues, this study proposes Transformer-Coordinate Attention-Efficient YOLO (TCE-YOLO), a detection framework designed with three key modules: (1) the Depthwise-Transformer-Vision (DTV) module integrates Depthwise Separable Convolutions (DSC), Vision Transformer, and multi-scale spatial pooling to efficiently represent local structures, spatial context, and global patterns of wheat seedlings; (2) the Feature Enhancement Module(FEM) incorporates coordinate attention to enhance seedling-related features while suppressing background interference; and (3) the Feature Coordination Module (FCM) performs multi-scale feature interaction with reduced computational cost. These components jointly improve robustness under dense planting and complex field conditions while maintaining lightweight deployment characteristics. Furthermore, we construct the Wheat Seedling Dataset (WSD), covering multiple planting densities, varieties, and field environments across two growing seasons. Experimental results show that TCE-YOLO outperforms mainstream detectors while maintaining high efficiency, providing a deployable solution for wheat seedling detection under real field conditions.
扫码关注我们
求助内容:
应助结果提醒方式:
