VM-YOLO: YOLO with VMamba for Strawberry Flowers Detection

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning

Background:

• Computer vision is crucial for non-invasive smart agriculture but faces challenges with limited computational resources on machinery.
• Optimizing algorithms for accuracy and computational power is vital for deploying advanced vision systems in agriculture.

Purpose of the Study:

• To develop a lightweight hybrid network, VM-YOLO, for efficient strawberry flower detection.
• To enhance the performance of object detection models on agricultural machinery with constrained computing power.

Main Methods:

• Proposed a novel lightweight hybrid network, VM-YOLO, integrating a Light C2f module and a VMambaNeck.
• Modified the YOLOv8 architecture by replacing the C2f module with Light C2f for multi-scale feature perception and the neck with VMambaNeck for global sensitivity.
• Trained and evaluated VM-YOLO on a custom strawberry flower dataset, performing ablation, multi-dataset, and comparative experiments.

Main Results:

• VM-YOLO demonstrated superior performance in object detection across various datasets compared to the baseline YOLOv8.
• The proposed VM-YOLO achieved better mean Average Precision (mAP), faster inference speed, and fewer parameters than YOLOv6, Faster R-CNN, FCOS, and RetinaNet.
• Ablation studies confirmed the effectiveness of the Light C2f module and VMambaNeck components.

Conclusions:

• VM-YOLO offers an effective solution for real-time strawberry flower detection in smart agriculture, addressing computational limitations.
• The hybrid network design provides a favorable trade-off between detection accuracy and computational efficiency, suitable for agricultural applications.
• This research contributes to advancing computer vision applications in precision agriculture by optimizing deep learning models for practical deployment.