MSRRT-DETR: A high-precision apple detection method with strong cross-domain generalization capability in complex...

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Accurate fruit detection is crucial for precision agriculture tasks like yield estimation and automated harvesting.
• Traditional models struggle with immature fruits, varietal differences, and complex orchard environments, leading to poor generalization and unstable predictions.

Purpose of the Study:

• To propose MSRRT-DETR, a novel fruit detection model balancing high accuracy, real-time performance, and robust generalization.
• To enhance the RT-DETR framework for complex agricultural scenarios.

Main Methods:

• Introduced a Multi-Scale Convolutional Attention Module (MSBlock) for improved multi-scale feature representation.
• Integrated a Spatial and Channel Synergistic Attention Module (SCSA) to boost object focus and discriminative ability.
• Implemented a Re-parameterized Feature Pyramid Network (RepGFPN) for efficient multi-scale feature fusion.

Main Results:

• MSRRT-DETR achieved 87.3% mAP50 on the TSApple dataset, outperforming YOLOv8, YOLO11, YOLO12, Faster R-CNN, Mask R-CNN, Cascade R-CNN, and RT-DETR variants.
• Inference speed reached 30.2 FPS, comparable to YOLO models, demonstrating a balance between accuracy and real-time capability.
• Showcased strong cross-domain generalization on public datasets like MinneApple, validating applicability across diverse scenarios and fruit varieties.

Conclusions:

• MSRRT-DETR effectively addresses limitations in current fruit detection models, offering high accuracy, fast inference, and strong generalization.
• The model provides robust technical support for intelligent monitoring and automated orchard management in precision agriculture.
• MSRRT-DETR holds significant practical value and broad application potential for complex agricultural scenarios.