基于EfficientNet的实时目标检测模型

REAL-TIME OBJECT DETECTION MODEL BASED ON EFFICIENTNET

在智能驾驶领域中,通常要求模型兼顾精度和推演速率。然而由于硬件条件的限制,目前大量的目标检测模型尚不能满足该要求。因此,基于单阶段目标检测算法提出一种准确率与推演速率相对平衡的实时目标检测模型。该模型使用EfficientNet-B1作为主干网络,SPP与改进后的PANet作为脖颈网络,CIoU损失函数与YOLO损失函数的组合作为模型训练时的损失函数。为了在不引入额外计算量的同时进一步提升模型的精度,在模型训练时引入多种模型训练技巧。实验结果证明,在BDD100K与PASCAL VOC数据集上,该模型相比YOLOv4模型有着更低的计算量和更好的检测精度,且实验中所使用的训练技巧均为模型带来了一定的精度提升,证明了该模型及训练技巧的有效性。

In the field of intelligent driving,most models require both accuracy and inference speed.However,due to the limitation of hardware,a large number of object detection models cannot meet the requirements.Therefore,based on the single-stage object detection algorithm,a real-time object detection model with relatively balance between accuracy and inference speed is proposed.The model used EfficientNet-B1 as the backbone,used SPP and improved PANet as the neck,and used the combination of CIoU loss and Yolo loss as the loss function during model training.In order to further improve the accuracy of the model without introducing additional computations,multiple model training tricks were introduced in model training.The experimental results show that the proposed model has better detection accuracy and lower computational complexity than YOLOv4 model on BDD100k and PASCAL VOC datasets,and the training tricks used in the experiment bring some accuracy improvement to the model,which proves the effectiveness of the model and training tricks.