摘要
针对当前传统网络模型对中药饮片检测精度低,检测不准确的问题,提出一种基于YOLOv8n优化改进的MSDA-YOLOv8中药饮片检测模型。首先,在Backbone上使用SCConv代替部分C2f模块,使用DyCAConv代替部分Conv。其次,添加DilateBlock模块,强化特征信息,提高了检测模型的特征融合能力。在Neck上,设计全新的C2fMSDA模块代替C2f,并引入Inception板块,扩大特征感受视野。使用BiFPN思想,高效双向跨尺度连接和加权特征融合,提高网络性能;最后将原有的损失函数替换为MPDIoU边界损失函数,模型的边界框回归性能有了提高。实验结果表明,改进后的YOLOv8模型在原模型的基础上提高识别精确度0.7%、平均精度2.9%,参数量降低1.9%。综合说明,该模型提高模型识别精度同时降低参数量,优于原算法以及对比算法,满足边缘计算要求,具有实际应用价值。
In response to the current issues of low accuracy and inaccurate detection in traditional network models for traditional Chinese medicines lices,this paper proposes an optimized and improved MSDA-YOLOv8 model for detecting Chinese medicine slices based on YOLOv8n.First,SCConv is used on the Backbone to replace some C2f module,and use DyCAConv replaces some Conv.Additionally,a DilateBlock module is added to enhance information in the features,which improves the model's feature fusion capability.On the Neck,a new C2fMSDA module is designed to replace C2f,introducing Inception block to expand the feature's receptive field.The BiFPN concept is employed for efficient bidirectional cross-scale connections and weighted feature fusion,improving network performance.Finally,the MPDIoU boundary loss function is used to replace basic loss function to improve the network's bounding box regression performance.Experimental results show that the improved model performs better in the dataset,with an increase of 0.7%in precision and 2.9%in mean average precision(mAP)compared to the original model.The model's parameter size is reduced by 1.9%compared to the original model.In summary,this model simultaneously reduces the model's parameter size while improving detection accuracy,significantly outperforming the comparison algorithms.It also meets the requirements of edge computing devices,demonstrating practical application value.
作者
华畅
郑豪
HUA Chang;ZHENG Hao(School of Artificial Intelligence and Information Technology,Nanjing University of Chinese Medicine/Jiangsu Prov-ince Engineering Research Center of TCM Intelligence Health Service,Nanjing 210023,China;School of Informa-tion and Engineering,Nanjing Xiaozhuang University,Nanjing 211171,China)
出处
《时珍国医国药》
CAS
CSCD
北大核心
2024年第12期2898-2904,共7页
Lishizhen Medicine and Materia Medica Research
基金
国家自然科学基金(61976118)。