基于Bottleneck的目标检测网络模型加速研究

Research on acceleration of object detection network model based on Bottleneck

针对YOLOv5的网络结构能降低模型复杂度,并解决数据不均衡问题,根据YOLOv5的结构特点,采取5种Bottleneck的替代方案,即IBN层、融合IBN层、Tucker卷积层、SPBottleneck和SEGBottleneck。实验基于COCO数据集,实验结果表明与改进前的Bottleneck相比,所提的IBN层、Tucker卷积层以及SEGBottleneck对模型的复杂度都有明显降低。模型规模分别减小了24.5%、22.5%和20.0%,模型运行速度分别提升了3.0%、3.8%和1.5%。基于Traffic数据集数据不均衡实验的结果表明:空间不均衡问题可以通过引入Focal EIoU解决;选择合适超参数能够加速模型的收敛,解决类别不均衡问题。

This paper aims at the network structure of YOLOv5 to reduce the complexity of the model and solve the problem of data imbalance.According to the structural characteristics of YOLOv5,five alternatives to Bottleneck were adopted,such as IBN layer,fusion IBN layer,Tucker convolutional layer,SPBottleneck and SEGBottleneck.This experiment was based on the COCO data set.The experimental results show that the proposed IBN layer,Tucker convolutional layer and SEGBottleneck significantly reduce the complexity of the model compared to the Bottleneck before improvement.The model size is reduced by 24.5%,22.5%,and 20.0%,and the model running speed is increased by 3.0%,3.8%,and 1.5%,respectively.The data imbalance experiment is based on the Traffic dataset.The experimental results of the IoU loss function show that the spatial imbalance problem can be solved by introducing Focal EIoU.The Focal Loss experiment shows that choosing appropriate hyperparameters can accelerate the convergence of the model and solve the problem of class imbalance.