| 摘 要: 针对胡麻果实因尺寸微小、分布密集及严重遮挡导致的检测难题，提出轻量化目标检测模型RGP-DETR。该模型基于GhostNet原理构建RGCSPELAN模块以降低约30%参数量；设计GLSA双分支注意力机制增强遮挡目标识别；新增P2检测层提升小目标检测效果。实验表明，模型大小仅12.1 MB，较基线模型RT-DETR-R18下降68.65%；FLOPs为29.5 G，较基线模型下降48.15%；mAP50达94.6%，召回率89%，实时性能达134.91 FPS。该研究展示了Transformer架构在农业小目标中的高效部署，为智慧农业提供高性价比视觉解决方案。 |
| 关键词: 目标检测 小目标检测 深度学习 胡麻植株果实 |
|
中图分类号:
文献标识码:
|
| 基金项目: 甘肃省自然科学基金 |
|
| Research on a lightweight detection model for flax plant fruits based on improved RT-DETR |
|
hongshicheng
|
Gansu Agricultural University
|
| Abstract: To address the detection challenges of flax fruits due to tiny size, dense distribution, and severe occlusion, a lightweight object detection model, RGP-DETR, is proposed. First, the RGCSPELAN module is constructed based on GhostNet principles to reduce parameters by approximately 30%. Second, a GLSA dual-branch attention mechanism is designed to enhance the recognition of occluded targets. Finally, an additional P2 detection layer is integrated to improve small object detection performance. Experimental results demonstrate that the model size is only 12.1 MB, a 68.65% reduction compared to the baseline RT-DETR-R18; FLOPs are 29.5 G, a 48.15% reduction; mAP50 reaches 94.6%, recall is 89%, and real-time performance achieves 134.91 FPS. This research demonstrates the efficient deployment of Transformer architecture for agricultural small objects, providing a cost-effective visual solution for smart agriculture. |
| Keywords: object detection small object detection deep learning flax plant fruits |
