| 摘 要: 针对库尔勒香梨成熟度人工检测效率低、主观性强且难以规模化的问题,提出了一种融合多场景数据构建、轻量化模型改进与移动端部署的一体化解决方案。首先,通过图像采集建立数据集,采用理化与感官评价双重标注方式,并结合场景增强与样本平衡提升数据质量。其次,改进 YOLOv8n模型:骨干网络嵌入 ADown细节提取、CGB背景区分与SPPF-LSKA多尺度融合模块;颈部引入GSConv分组卷积;进一步通过INT8量化与通道剪枝实现轻量化,并优化CPU训练参数。最后,开发了支持实时分析的移动端检测系统。实验表明,改进模型 Top-1准确率达90.89%,模型体积压缩至8.5MB,CPU推理速度为0.58s/张。移动端系统在果园与市场场景准确率分别为89.5%与88.8%。通过“数据到模型再到系统”协同优化,为特色农产品智能检测提供了可落地的轻量化方案。 |
| 关键词: 库尔勒香梨 YOLOv8 成熟度检测 轻量化模型 移动端部署 |
|
中图分类号:
文献标识码: A
|
|
| Maturity Detection System for Korla Fragrant Pears Based on Improved YOLOv8 |
|
PENG Lingying, WU Naining, FAN Shuchen
|
(College of Computer and Information Engineering, Xinjiang Agricultural University, Urumqi 830052, China)
3257555949@qq.com; 66642599@qq.com; 3280633995@qq.com
|
| Abstract: Aiming at the problems of low efficiency, strong subjectivity, and difficulty in large-scale application in manual maturity detection for Korla fragrant pears, this paper proposes an integrated solution combining mult-i scenario data construction, lightweight model improvement, and mobile deployment. Firstly, images are collected to establish a dataset with dual labeling based on physicochemical and sensory evaluation, and data quality is improved through scene enhancement and sample balancing. Secondly, the YOLOv8n model is improved: the ADown detail extraction module,CGB background discrimination module, and SPPF-LSKA multi-scale fusion module are embedded into the backbone network, while GSConv grouped convolution is introduced into the neck network. Furthermore, model lightweighting is achieved via INT8 quantization and channel pruning, and CPU training parameters are optimized. Finally, a mobile detection system supporting rea-l time analysis is developed. Experimental results demonstrate that the improved model achieves a Top-1 accuracy of 90.89% , with the model size compressed to 8.5 MB and a CPU training speed of 0.58 seconds per image. The mobile system achieves accuracies of 89.5% and 88.8% in orchard and market scenarios,respectively. Through collaborative optimization from data to model and then to system, this study provides a deployable lightweight scheme for intelligent detection of characteristic agricultural products. |
| Keywords: Korla fragrant pear YOLOv8 maturity detection lightweight model mobile deployment |
