摘要
针对原始YOLOv8n算法小目标零件识别率低的问题,在原算法的基础上添加CBAM注意力机制,从不同的维度学习图像特征,引入基于动态非单调聚焦机制的边界框损失函数,定义一个离群度表示预测框的质量,在回归质量较好和回归质量较差的样本中作出平衡,充分挖掘非单调聚焦机制的潜能。在自制的小目标零件数据集上,分别基于原始YOLOv8n和改进后的YOLOv8n算法进行实验。实验结果表明,改进后的YOLOv8n模型零件识别准确率约为85.8%,平均准确率提升6.8%。
This paper addesses the issue of low recognition rate of small target parts of the original YOLOv8n algorithm,CBAM attention mechanism is added to learn image features from different dimensions,a bounding box loss function based on dynamic non-monotonic focusing mechanism is introduced,and an outlier is defined to represent the quality of prediction box,to make a balance between the samples with better regression quality and those with poorer regression quality,and to fully exploit the potential of the non-monotonic focusing mechanism′s potential.Experiments based on the original YOLOv8n and the improved YOLOv8n algorithms are conducted on a home-made small target part dataset,respectively.The experimental results show that the improved YOLOv8n model part recognition accuracy is about 85.8%with an average accuracy improvement of 6.8%.
作者
王钧
麻方达
符朝兴
WANG Jun;MA Fangda;FU Chaoxing(College of Mechanical and Electrical Engineering,Qingdao University,Qingdao 266071,China)
出处
《青岛大学学报（工程技术版）》
CAS
2024年第2期39-46,共8页
Journal of Qingdao University(Engineering & Technology Edition)
基金
山东省自然科学基金资助项目(ZR2020QE183)。
