【作者】 袁旭东；

【导师】 曲海成；

【作者基本信息】 辽宁工程技术大学 ， 电子信息硕士（专业学位）， 2023， 硕士

【摘要】 佩戴口罩是常态化疫情防控的要求,利用边缘设备实现口罩佩戴检测可以减少人工成本。为了让目标检测模型更好地适用于边缘设备,从YOLOX-Tiny的模型大小和检测精度两个方面进行改进,提出了Lite-YOLOX。为了降低原模型大小并提高检测速度,重新设计了特征金字塔的结构,首先增加金字塔特征的输入,然后对输入的通道进行统一,最后调节特征金字塔的结构,使特征可以更有效地进行融合,进而形成了一个高效且轻量的特征融合结构;对于模型的检测头部分,在保持感受野不变的前提下,减少了卷积层的使用,同时使用深度可分离卷积替代普通卷积。为了提高模型的检测精度,对原有的Io U损失函数进行优化,设计并提出EIo U损失函数,改进后的损失函数对真实框和预测框的位置更加敏感;模型的标签匹配算法存在真实框没有匹配正样本缺陷,对该算法进行了优化,使每个真实框都有与其匹配的正样本。实验中选取PASCAL VOC、AICOO和自制数据集对改进模型效果进行验证。实验结果表明:Lite-YOLOX相比YOLOX-Tiny,参数量减少40%,计算量下降37.5%;在PASCAL VOC数据集上,检测精度提高了4.0%;在AIZOO数据集上,检测精度提高了2.8%;在自制数据集上,检测精度提高了3.8%。实际部署在NVIDIA Jetson Xavier NX上,FPS从51增加到59。该论文有图30幅,表12个,参考文献58篇。更多还原

【Abstract】 Wearing masks is a requirement for normalized epidemic prevention and control,and edge devices are also used to detect mask wearing to reduce labor costs.To this end,we need to improve from both the model size and detection accuracy of YOLOX-Tiny to make the object detection model better suited for edge devices.To compress the size of the model and improve the detection speed,the structure of the feature pyramid is redesigned.First,the input of pyramid features is increased,then the input channels are unified,and finally the structure of the feature pyramid is adjusted so that features can be fused more effectively,forming an efficient and lightweight feature fusion structure.For the detector head of the model,while keeping the receptive field unchanged,the use of convolution layer is reduced,and the depth separable convolution is used to replace the ordinary convolution.In order to improve the detection accuracy of the model,the original Io U loss function is optimized,and the EIo U loss function is designed and proposed.The improved loss function is more sensitive to the position of the real box and the prediction box.The label matching algorithm of the model has defects.The algorithm is improved so that every real box has its matching sample.PASCAL VOC,AICOO and self-made datasets are selected to verify the improved model.The experimental results show that Lite YOLOX has 40% less parameters and37.5% less computation than YOLOX Tiny.On PASCAL VOC dataset,the detection accuracy is improved by 4.0%;On the AIZOO dataset,the detection accuracy is improved by 2.8%;On the self-made data set,the detection accuracy is improved by 3.8%.In addition,the model is deployed and tested on an NVIDIA Jetson Xavier NX device,and the FPS indicator is increased from 51 to59.There are 30 figures,12 tables and 58 references in this thesis.更多还原