基于改进YOLO v8s的自然环境下荔枝果实检测方法

Method for Detecting Litchi Fruits in Natural Environment Based on Lightweight Improved YOLO v8s

摘要:

随着智慧农业技术的推广，针对自然环境下荔枝果实目标尺寸小、遮挡现象严重与背景复杂导致的检测精度较低以及实时检测设备的资源限制等问题，提出了基于改进YOLO v8s的轻量化荔枝果实检测方法。采用轻量级网络MobileNetV3作为主干网络，减小模型参数量;引入卷积局部注意力模块（Convolutional local attention module，CLAM），同时在通道和空间上提高模型在自然环境下对荔枝果实特征提取能力，并引入残差学习理念，将通过注意力模块前后的特征进行加权融合，保证了模型对原始图像的特征学习，提高模型在复杂环境下的检测稳定性;将部分卷积层替换为深度可分离卷积，并进行多尺度特征融合;采用αEIoU作为损失函数，提高目标检测边界框纵横比收敛速度，减小重叠果实漏检率。试验结果表明，改进YOLO v8s在试验数据集上精确率和平均精度分别达到91.75%和79.07%，相较原始模型分别提升17.29、14.75个百分点，同时参数量减至5.488×10?，较原始模型降低50.7%。与主流轻量级模型YOLO v5s、EfficientNetV2、YOLO v7-Tiny、YOLO v9s、YOLO v10s与YOLO 11s相比，本文模型在精确率、召回率和平均精度均更优，为现代果园环境下移动设备的荔枝检测及测产提供了技术参考。

Abstract:

With the promotion of smart agriculture technology, there are issues such as low detection accuracy and resource limitations of real-time detection devices due to the small size of litchi fruits, severe occlusion, and complex background in natural environments. A lightweight litchi fruit detection method was proposed based on the improved YOLO v8s. Firstly, a lightweight network, MobileNetV3, was adopted as the backbone network to reduce the model parameters. On this basis, the convolutional local attention module (CLAM) was introduced to enhance the model's feature extraction ability for litchi fruits in natural environments both in channels and space. The concept of residual learning was also introduced to fuse the features before and after the attention module through weighted addition, ensuring the model's feature learning from the original image and improving the detection stability in complex environments. Secondly, some convolutional layers were replaced with depthwise separable convolutions, and multi-scale feature fusion was performed. Finally, to address the issues of small fruit size and severe occlusion, the αEIoU was used as the loss function to accelerate the convergence speed of the aspect ratio of the detection bounding box and reduce the missed detection rate of overlapping fruits. Experimental results indicated that the improved YOLO v8s achieved an accuracy rate of 91.75% and a detection precision of 79.07% on the experimental dataset, which were 17.29 and 14.75 percentage points higher than that of the original model, respectively. At the same time, the number of parameters was significantly reduced to 5.488×10?, a decrease of 50.7% compared with that of the original model. Compared with mainstream lightweight models such as YOLO v5s, EfficientNetV2, YOLO v7-Tiny, YOLO v9s, YOLO v10s, and YOLO 11s, this model demonstrated advantages in terms of precision, recall, and detection accuracy, providing a technical reference for litchi detection and yield estimation in modern orchard environments using mobile devices.