基于FPBW-YOLO v8的复杂场景下番茄果实识别方法

Tomato Fruit Recognition in Complex Scenes Based on FPBW-YOLO v8

摘要:

番茄果实的快速准确检测是实现其智能采摘的重要前提，针对部署需求以及番茄图像背景复杂、枝叶遮挡和果实重叠等问题，提出了一种基于改进YOLO v8n的复杂场景下番茄果实识别方法。使用FasterNet作为YOLO v8n的主干特征提取网络，提高模型的特征提取能力；通过在颈部网络中融合P2小目标检测分支和双向特征金字塔网络（Bidirectional feature pyramid network，BiFPN）结构，降低复杂背景的干扰以提高模型的检测精度；使用WIoU（Wise intersection over union）损失函数提高模型在遮挡、重叠情况下的果实定位精度，同时加快模型收敛；基于NCNN框架将模型部署到移动端进行测试。试验结果表明，FPBW-YOLO v8模型精确率、召回率、平均精度均值（mAP@0.5和mAP@0.5:0.95）分别达到97.9%、95.1%、98.3%和74.3%，相较于Faster R-CNN、SSD、YOLO v8n、YOLO v7、YOLO v5n和Rt-Detr均有明显优势，且模型内存占用量仅87MB。在移动端上的测试结果表明，本文模型能够在计算资源有限的硬件设备上获得较高的检测精度，可以有效解决复杂场景下番茄果实的识别问题，为番茄采摘机器人的研发提供技术支持。

Abstract:

Rapid and accurate detection of tomato fruits is an important prerequisite of intelligent harvesting. Aiming at deployment requirements and problems of complex background, branch and leaf occlusion, and overlapping of fruits, an improved detection method based on improved YOLO v8n was proposed. Firstly, FasterNet was selected as the backbone feature extraction network to improve the feature extraction capability of the model. Nextly, the small target detection layer and bidirectional feature pyramid network (BiFPN) structure were fused in the neck network, reducing the interference of the complex background to improve the detection accuracy of the model. Subsequently, the wise intersection over union (WIoU) loss function was used to improve the detection accuracy of the model in the presence of occlusion and overlap, enhancing the model’s convergence ability. Finally, the model was deployed to a mobile terminal based on the NCNN framework. Experimental results on the tomato public dataset showed that the precision, recall, mAP@0.5 and mAP@0.5:0.95 of the FPBW- YOLO v8 model reached 97.9%, 95.1%, 98.3% and 74.3%, respectively, all of which were higher than the results of Faster R-CNN, SSD, YOLO v8n, YOLO v7, YOLO v5n and Rt-Detr. The model brought forward in this research could obtain high detection accuracy on hardware devices with limited computational resources, which can effectively solve the problem of tomato fruit recognition in complex scenes and provide technical support for tomato picking robots.