SDA-YOLO: An Object Detection Method for Peach Fruits in Complex Orchard Environments

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Peach detection in orchards faces challenges like occlusion, complex backgrounds, and scale variations.
• Existing methods struggle with accurately identifying peaches under these difficult conditions.

Purpose of the Study:

• To develop an improved peach detection method, SDA-YOLO, based on YOLOv11n for complex orchard environments.
• To enhance feature representation, localization accuracy, and feature fusion flexibility for peach detection.

Main Methods:

• Integrated LSKA module into SPPF for multi-scale feature representation (SPPF-LSKA).
• Employed MPDIoU loss for improved bounding box regression of occluded peaches.
• Incorporated DyHead Block into detection head (DMDetect) for better feature discrimination.
• Introduced Adaptive Multi-Scale Fusion Pyramid (AMFP) module to enhance neck network flexibility.

Main Results:

• SDA-YOLO achieved precision of 90.8%, recall of 85.4%, mAP@0.95 of 90%, and mAP@0.5:0.95 of 62.7%.
• Demonstrated significant improvements over the baseline YOLOv11n, with increases of 2.7%, 4.8%, 2.7%, and 7.2% respectively.
• Verified robustness in complex orchard settings.

Conclusions:

• SDA-YOLO offers a robust solution for peach detection in challenging orchard environments.
• The method provides effective technical support for intelligent fruit harvesting and yield estimation.