【摘要】 为了实现多种水果在采摘后自动化筛选和分拣中腐败水果识别的问题,提出了改进的YOLOv5n模型,命名为mobile-YOLO。首先将YOLOv5n的主干网络替换为MobileNetV3并引入深度可分离卷积,相较于原模型,这种改进在计算效率和速度上都有所提升,并且准确率也得到了提高。为了进一步提升速度,将C3模块替换为C2f模块,实现轻量化的同时获得了更丰富的梯度流信息。最后将原有的CIoU替换为α-CIoU,以加快收敛速度并保证图像框位置的准确性。mobile-YOLO相较于原始的YOLOv5n,mAP@.5 (mean Average Precision)达到了98.1%,mAP@.5:.95达到了94.2%,同时在P(Precision)值为97.1%和R(Recall)值为96.8%的情况下,参数量几乎与YOLOv5n保持一致。此外,计算量下降至3.7 GFLO/s,实现了实时高精度的腐败水果检测。同时,在CPU、NVIDIA Jetson Nano和NVIDIA Jetson Xavier NX等设备上部署测试,帧率检测结果 分别为20 F/s、3 8 F/s和76 F/s,满足在边缘设备上进行实时检测的需求,验证了提出的mobile-YOLO模型在腐败水果分拣识别方面具有较强的实用性。更多还原

【Abstract】 To address the issues of automated sorting and classification of various fruits post-harvest, particularly in the identification of decayed fruits, this paper introduces an enhanced model derived from YOLOv5n, denominated as mobile-YOLO. The initial modification involves replacing the backbone network of YOLOv5n with MobileNetV3 and incorporating depth-wise separable convolutions. This enhancement results in improved computational efficiency and speed compared to the original model, accompanied by an elevation in accuracy. To further enhance speed, the C3 module is substituted with the C2f module, achieving lightweight design while acquiring richer gradient flow information. Additionally, the original CIoU loss is substituted with α-CIoU to expedite convergence and ensure precise localization of bounding boxes. Mobile-YOLO, compared to the original YOLOv5n, attains a mAP@.5 of 98.1%, a mAP@.5:.95 of 94.2%, and maintains parameter quantity nearly consistent with YOLOv5n under P-value(Precision) of 97.1% and R-value(Recall) of 96.8%. Furthermore, the computational load is reduced to 3.7 GFLO/s, enabling real-time high-precision detection of decayed fruits. Deployment tests on CPU, NVIDIA Jetson Nano, and NVIDIA Jetson Xavier NX devices yield detection frame rates of 20 F/s, 38 F/s, and 76 F/s, respectively. This confirms the robust practicality of the proposed mobile-YOLO model for decayed fruit sorting and recognition on edge devices, meeting the real-time detection requirements.更多还原