YOLO-CDD: Real-time and accurate detection of latex collection bowl and flow states with enhanced multi-scale feature fusion

Natural rubber is an essential industrial material extensively utilized in aerospace, transportation, and healthcare. Nonetheless, latex harvesting continues to depend significantly on manual assessment of collection bowl statuses (full, not full, tipped over) and flow trail conditions (exhausted, not exhausted), leading to inefficiency and elevated labor expenses—thus falling short of the requirements of contemporary precision agriculture. To effectively identify latex collection bowls and flow states in natural rubber harvesting, deep learning offers a scalable and intelligent solution for real-time monitoring. This paper presents an improved object detection model based on YOLOv12n, named YOLO with C2f-DSC and DWConv enhancements (YOLO-CDD), specifically designed for high-precision and real-time detection of latex collection bowls and latex flow states. An important innovation is the implementation of a self-designed Cross-Stage Partial Fusion with Depthwise Separable Convolution (C2f-DSC), which integrates multi-scale depthwise separable convolutions with an augmented C2f architecture to enhance multi-scale feature extraction and representation. This improves the model's capacity to identify objects at different scales. Furthermore, lightweight Depthwise Separable Convolution (DWConv) is incorporated into both the backbone and head network to diminish computational expense and parameter quantity without sacrificing accuracy, thereby enhancing the model's applicability for real-world implementation. A specialized dataset comprising latex collection bowls and flow trails was created to establish a robust foundation for this research. Experimental findings on the dataset indicate that YOLO-CDD attains a mAP0.50 of 80.7%, exceeding the baseline YOLOv12n by 3.6%, and enhancing mAP0.50:0.95 by 3.7%. Moreover, inference time and GFLOPs are diminished by 13.3% and 3.2%, respectively. In comparison to other advanced models like YOLOv10n and YOLOv11n, YOLO-CDD also exhibits significant enhancements across essential metrics. The YOLO-CDD model enhances detection accuracy while preserving high computational efficiency, rendering it ideal for real-time monitoring tasks. • This paper proposes a multi-scale convolution and feature-enhanced module (C2f-DSC), which guides the model to more effectively extract key features at different scales, enhancing robustness under complex backgrounds and reducing false and missed detections in rubber harvesting. • The DWConv is introduced and integrated with the C2f-DSC at key positions in the backbone, reducing parameters and computation while enhancing detection performance, to support efficient lightweight deployment in complex environments. • A dataset comprising 2,408 annotated images was constructed, capturing diverse latex collection bowl and flow trail states under realistic rubber plantation conditions. • YOLO-CDD improves mAP0.50 and mAP0.50:0.95 by 3.6% and 3.7%, respectively, compared to the original YOLOv12n model, while reducing inference time and GFLOPs by 13.3% and 3.2%. It also outperforms other advanced models such as YOLOv10n and YOLOv11n across key metrics.