CR-YOLOv9: Improved YOLOv9 Multi-Stage Strawberry Fruit Maturity Detection Application Integrated with CRNET.

Abstract

Strawberries are a commonly used agricultural product in the food industry. In the traditional production model, labor costs are high, and extensive picking techniques can result in food safety issues, like poor taste and fruit rot. In response to the existing challenges of low detection accuracy and slow detection speed in the assessment of strawberry fruit maturity in orchards, a CR-YOLOv9 multi-stage method for strawberry fruit maturity detection was introduced. The composite thinning network, CRNet, is utilized for target fusion, employing multi-branch blocks to enhance images by restoring high-frequency details. To address the issue of low computational efficiency in the multi-head self-attention (MHSA) model due to redundant attention heads, the design concept of CGA is introduced. This concept aligns input feature grouping with the number of attention heads, offering the distinct segmentation of complete features for each attention head, thereby reducing computational redundancy. A hybrid operator, ACmix, is proposed to enhance the efficiency of image classification and target detection. Additionally, the Inner-IoU concept, in conjunction with Shape-IoU, is introduced to replace the original loss function, thereby enhancing the accuracy of detecting small targets in complex scenes. The experimental results demonstrate that CR-YOLOv9 achieves a precision rate of 97.52%, a recall rate of 95.34%, and an mAP@50 of 97.95%. These values are notably higher than those of YOLOv9 by 4.2%, 5.07%, and 3.34%. Furthermore, the detection speed of CR-YOLOv9 is 84, making it suitable for the real-time detection of strawberry ripeness in orchards. The results demonstrate that the CR-YOLOv9 algorithm discussed in this study exhibits high detection accuracy and rapid detection speed. This enables more efficient and automated strawberry picking, meeting the public's requirements for food safety.

Plant phenotyping relevance

イチゴ果実の成熟度という植物形質を画像から検出する手法を開発し、精度・再現率・mAP・速度で評価しており、フェノタイピング手法が研究の中心である。

Code and data availability

The paper describes a custom strawberry maturity image dataset (RealSense D455 captures of 'Red Face' strawberries, LabelImg-annotated) and the CR-YOLOv9 model, but provides no public repository, deposit, or URL for the dataset, code, or trained model. The Data Availability Statement only offers contact with the对应ing作者

No evidence-backed public reproduction asset is currently recorded.