改进SSD算法及其在口罩佩戴检测中的应用

Improved SSD Algorithm and Its Application in Mask Wear Detection

SSD目标检测算法虽然具有检测速度快且检测精度较高的优点,但由于采用smoothL1损失函数作为定位损失函数,存在着边界框回归不准确的问题。针对上述问题提出了一种基于SSD改进的目标检测算法,设计了EIoUloss作为新的定位损失函数,实现了目标函数最优即为IoU局部最优。实验结果表明,改进SSD算法通过修改定位损失函数,可以提高目标检测中预测框的准确性,改进算法较原SSD算法在VOC2007数据集和口罩佩戴数据集上mAP-50分别提高了1.97%和1.96%。

Although the SSD object detection algorithm has the advantages of fast detection speed and high detection accuracy,there is a problem of inaccurate bounding box regression because the smoothL1 loss function is used as the positioning loss function.Aiming at this problem,an improved object detection algorithm based on SSD is proposed.EIoUloss was designed as a new localization loss function to achieve that the optimal object function was the optimal local IoU.Experimental results show that the improved SSD algorithm can improve the accuracy of the object detection prediction box by modifying the function of the localization loss.Compared with the original SSD algorithm,the mAP-50 on the VOC2007 data set and mask wearing data set is increased by 1.97% and 1.96% respectively.