可堆叠重校准特征金字塔目标检测方法

无人机、车载识别等边缘设备应用日益增长，对模型参数量、检测速度以及精度提出了进一步的要求。为了提高目标检测在这些领域的项目落地能力，提出一种可堆叠重校准特征金字塔模块以及改进的SR-YOLOv3目标检测网络，使用对边缘设备友好的主干网络作为特征提取网络，通过堆叠轻量的金字塔模块，在减少参数数量的同时，提高检测精度及速度。在公开的目标检测数据集PascalVOC上进行性能评估，实验结果显示，该改进算法的参数量有明显下降，且计算速度得到提升。

Object detection based on stackable recalibrated feature pyramid networks

With the increasing application of edge devices such as unmanned aerial vehicles and vehicle identification, further requirements are put forward on the amount of model parameters, detection speed and accuracy. Therefore, the paper proposes a stackable recalibrated feature pyramid network and an improved SR-YOLOv3 object detection network, using a backbone network that is friendly to edge devices as a feature extraction network. The lightweight pyramid module reduces the number of parameters while increasing the detection accuracy and speed. The experimental results show that based on the performance evaluation on the public target detection dataset PascalVOC, the parameter amount of the improved algorithm has dropped significantly, and the calculation speed has also been improved.