摘要
近年来,基于深度学习的目标检测算法发展极为迅速。YOLOv8s能够兼顾检测精度和速度,有利于脐橙智能化采摘的实现。本文分别从数据集扩充和YOLOv8s模型改进两方面提高脐橙的检测精度。数据集扩充方面,对晴天采集的脐橙图像进行加雨和加雾扩充,提高模型在复杂环境下的检测能力。模型改进方面,先将YOLOv8s颈部的特征融合改进为权重化的融合,以突出重要的特征;然后将YOLOv8s原浅层处的检测头更换至更浅层,以检测出因遮挡严重而表现为小的目标;最后,将YOLOv8s主干网络的第二个卷积改进为全维度动态卷积,以在卷积的空间、输入通道、输出通道和整个卷积核四个维度上关注重要特征。实验结果表明,与未对数据集扩充且采用原始YOLOv8s模型的方法相比,本文方法获得的精确率、召回率和平均精确率均值都得到大幅度提升。
In recent years, deep learning-based object detection algorithms have developed rapidly. YOLOv8s can balance detection accuracy and speed, which is conducive to the implementation of intelligent navel orange picking. In this paper, we improve the detection performance from two aspects: dataset expansion and model improvement. In terms of dataset expansion, rain and fog are added into the navel orange images collected on sunny days to improve the detection ability of the model in complex environments. In terms of model improvement, the weighed concatenation operation is adopted in feature fusion to highlight the important features;the detection head at the shallow layer in the original model is transferred to a shallower layer to detect targets that appear small due to the severe occlusion;the second convolution operation in the backbone is replaced by a omni-dimensional convolution to focus on important features in four dimensions: spatial position, input channel, output channel and the entire convolution kernel. The experimental results show that when compared with the method using the original YOLOv8s without data augmentation, the precision, the recall and the mean average precision obtained by our method have been significantly improved.
出处
《计算机科学与应用》
2024年第6期41-49,共9页
Computer Science and Application
