YOLO-DCS: A high-precision edge detection method for latex bowl state and latex trail state

The advancement of smart agriculture has rendered the replacement of conventional manual harvesting techniques with automated harvesting systems an unavoidable trend. Attaining swift and accurate identification of the latex bowl’s condition and the latex trail presents a considerable challenge. To achieve edge-optimized precision detection, YOLO+DFA+DWConv+C2f-CPCA+SPDConv(YOLO-DCS), an edge-capable detection model based on the YOLOv10s framework, is proposed for the swift and accurate identification of latex collection bowl states (full, not full, tipped over) and latex trail states (run out, not run out). YOLO-DCS introduces a novel Dynamic Feature Aggregation (DFA) module designed to consolidate dynamic contextual information to improve feature fusion. It allows the model to more precisely discern nuanced variations in the condition of collection bowls, including differences in texture and color. A refined C2f module integrating CPCA channel attention (C2f-CPCA) is introduced to augment the representation of essential features, thereby substantially enhancing the model’s efficacy in small object detection and edge computing contexts. The integration of lightweight Depthwise Convolution (DWConv) and SPDConv into both the backbone and head networks significantly diminishes the model’s parameter size while preserving high recognition accuracy, resulting in a notable enhancement in inference speed. Experiments conducted on our self-collected visible-light dataset demonstrate that the proposed YOLO-DCS model attains state-of-the-art performance. The model achieves a mAP0.50 of 82.1% and a mAP0.50:0.95 of 70.3%, reflecting enhancements of 4.7% and 2.2% compared to the original YOLOv10s model, respectively. The recall and F1 Score attain 78.5% and 84.2%, exceeding the original YOLOv10s by 9.3% and 8.0%, respectively. Significantly, YOLO-DCS attains a GFLOPs of merely 8.6, representing a 64.9% decrease from the 24.5 GFLOPs of the YOLOv10s model, underscoring its remarkable lightweight efficiency. Moreover, supplementary comparative experiments were conducted on a self-collected infrared dataset to further validate the generalization and robustness of the YOLO-DCS model. The experimental results show that YOLO-DCS still achieves improvements of 3.3% in mAP0.50 and 3.5% in recall, while significantly reducing GFLOPs by 64.9% compared to the original YOLOv10s model. The findings indicate that YOLO-DCS not only enhances the automation and accuracy of rubber harvesting in smart agriculture but also exhibits strong adaptability and robustness across both visible and infrared scenarios, highlighting its potential for deployment in real-world applications. • A self-designed DFA module dynamically refines features via adaptive convolution, boosting robustness. • The novel C2f-CPCA module integrates C2f and CPCA attention to better capture essential small targets. • The lightweight DWConv and SPDConv modules reduce parameters and size while maintaining multi-scale learning ability. • The YOLO-DCS model improves mAP by 4.7%, recall by 9.3%, and F1 Score by 8.0%, while reducing GFLOPs by 64.9%.