An Enhanced YOLO Framework for Small Object Detection in Complex Agricultural Environments

Abstract To address challenges such as small-object detection, dense occlusion, and background interference in complex greenhouse environments within intelligent agriculture, this study proposes an enhanced YOLOv7-Rcs object detection model. Specifically, the model utilizes an ELAN-Rep backbone integrating RepVGG modules to improve feature representation with reduced computational cost, introduces a Channel-Spatial Attention Mechanism (CSAM) to refine salient features and suppress noise, and incorporates a multi-task small-object detection head to enhance fine-scale target recognition. Experimental results indicate that this model significantly outperforms the standard YOLOv7 on WiderPerson, Open Images V6, and a custom greenhouse dataset, achieving up to 90.9% mAP while maintaining strong real-time performance. Therefore, the YOLOv7-Rcs model demonstrates superior visual detection accuracy and robustness in real greenhouse environments, effectively supporting precision agriculture applications such as automated crop monitoring and greenhouse inspections.