A Lightweight Model for Shine Muscat Grape Detection in Complex Environments Based on the YOLOv8 Architecture

Chunna Tian, Zhepeng Liu, Haosen Chen, Fanglong Dong, Xixiang Liu, Cong LinAgronomy 2025 · DOI

TL;DR

This paper introduces a lightweight YOLOv8-based model for detecting and classifying "Sunshine Rose" grapes in challenging orchard conditions. By incorporating novel modules like DualConv and C2f-GND, the model achieves improved accuracy and significantly reduces computational cost and inference time, making it suitable for deployment on resource-constrained edge devices for automated harvesting.

What Problem Does This Paper Address?

Automated grape harvesting, particularly for varieties like "Sunshine Rose," faces significant hurdles in complex vineyard environments. These challenges include variable lighting, occlusion by leaves and other clusters, and the need to distinguish between healthy and damaged grapes. Current detection models, while improving, often require substantial computational resources, making them unsuitable for deployment on edge devices commonly used in agricultural robotics. This limits the practical application of advanced AI for tasks like automated sorting and harvesting, which are crucial for addressing labor shortages and improving efficiency in the grape industry.

Methods

The core of this research is a modified YOLOv8 architecture. The authors introduce several key innovations:

  • DualConv: This module replaces standard convolutions with a combination of Group Convolution and heterogeneous convolution. It aims to enhance feature extraction and information fusion while reducing computational cost and parameter count.
  • C2f-GND Module: This novel module replaces the standard C2f module in YOLOv8. It incorporates the Ghost Module to reduce parameters and FLOPs by generating feature maps through efficient linear operations and depthwise convolutions. Additionally, it integrates Dynamic Convolution, which allows the model to adaptively select and combine convolutional kernels ("experts") to process input data, enhancing adaptability to varying conditions without significantly increasing FLOPs.
  • Lightweight Design: The overall goal is to create a model that is both accurate and computationally efficient, suitable for edge deployment.

Datasets

  • Shine-Muscat-Complex Dataset: A newly constructed dataset specifically for this study, containing 4715 images of "Sunshine Rose" grapes. This dataset is designed to capture diverse real-world scenarios, including varying lighting, occlusion, and different angles. It classifies grapes into two categories: "Standard Grapes" and "Compromised Grapes."
  • Existing Datasets: While not explicitly used for training the final model, the paper references other grape detection datasets (e.g., GrapeCS-ML) in its literature review.

Key Insights and Ideas

  • Balancing Act: The most significant insight is the successful balancing of accuracy and computational efficiency. The authors demonstrate that by strategically modifying existing modules and introducing new ones, it's possible to achieve state-of-the-art performance while drastically reducing the model's footprint.
  • Ghost Module's Efficiency: The Ghost Module proves effective in reducing parameters and FLOPs by cleverly generating feature maps, showing that not all feature extraction needs to be computationally intensive.
  • Dynamic Convolution for Adaptability: The integration of Dynamic Convolution is crucial for handling the variability in orchard environments. It allows the model to adapt its processing based on the input, which is key for robustness against lighting changes and occlusions.
  • C2f-GND's Synergy: The C2f-GND module, by combining the benefits of Ghost Module and Dynamic Convolution, offers a powerful enhancement to the YOLOv8 backbone, leading to improved feature extraction and detection capabilities, especially for small and overlapping targets.

Key Results

  • Improved mAP: The proposed model achieved a 2.6% improvement in mean Average Precision (mAP) over the baseline YOLOv8n.
  • Significant Reduction in Size and Speed: The model boasts a 36.8% reduction in parameters and a 34.1% reduction in FLOPs compared to YOLOv8n. Crucially, inference time was reduced by 15%.
  • Competitive Performance: Compared to YOLOv11n, the proposed model achieved a 3.3% higher mAP with substantial reductions in parameters (26.4%), FLOPs (14.3%), and inference time (14.6%).
  • Robustness: The model demonstrated superior performance in detecting small and overlapping targets, as well as in diverse lighting conditions, outperforming YOLOv8n in qualitative evaluations.

Limitations and Future Work

  • Dataset Specificity: The study's reliance on the custom Shine-Muscat-Complex dataset might lead to overfitting. The authors suggest that further validation on other datasets and crop types is necessary to assess generalizability.
  • Classification Nuances: While detection is strong, the classification accuracy between "Standard Grapes" and "Compromised Grapes" could be further improved. Fine-tuning architecture or exploring advanced feature extraction techniques are suggested.
  • Broader Applicability: Adapting the model to other fruits and vegetables will require modifications to the architecture and training strategies, necessitating testing on heterogeneous datasets.
  • Future Directions: The authors plan to further investigate edge deployment, expand data collection to more grape varieties and crops, and explore transfer learning to enhance robustness and utility across diverse agricultural applications.

Summary updated

Abstract

Automated harvesting of “Sunshine Rose” grapes requires accurate detection and classification of grape clusters under challenging orchard conditions, such as occlusion and variable lighting, while ensuring that the model can be deployed on resource- and computation-constrained edge devices. This study addresses these challenges by proposing a lightweight YOLOv8-based model, incorporating DualConv and the novel C2f-GND module to enhance feature extraction and reduce computational complexity. Evaluated on the newly developed Shine-Muscat-Complex dataset of 4715 images, the proposed model achieved a 2.6% improvement in mean Average Precision (mAP) over YOLOv8n while reducing parameters by 36.8%, FLOPs by 34.1%, and inference time by 15%. Compared with the latest YOLOv11n, our model achieved a 3.3% improvement in mAP, with reductions of 26.4% in parameters, 14.3% in FLOPs, and 14.6% in inference time, demonstrating comprehensive enhancements. These results highlight the potential of our model for accurate and efficient deployment on resource-constrained edge devices, providing an algorithmic foundation for the automated harvesting of “Sunshine Rose” grapes.
BibTeX
@inproceedings{tian2025lightweight,
  author = {Chunna Tian and Zhepeng Liu and Haosen Chen and Fanglong Dong and Xixiang Liu and Cong Lin},
  title = {{A Lightweight Model for Shine Muscat Grape Detection in Complex Environments Based on the YOLOv8 Architecture}},
  booktitle = {Agronomy},
  year = {2025},
  doi = {10.3390/agronomy15010174},
}
Figure