基于注意力机制和尺度均衡金字塔网络的目标检测

Object detection based on attention mechanism and scale equalization pyramid network

在目标检测领域中,基于预先是否对预测框与标注框进行关联,可分为anchor based与anchor free两种方法.Anchor based多应用于早期目标检测方法中,能够提升检测准确率,但过程复杂且计算量较大,因此会增加模型训练时间与推理速度,从而导致应用效率大大降低.而基于anchor free的方法去除了关于anchor的冗余计算,提升模型推理速度,但同时也牺牲了一定的准确率.结合二者的优点并基于anchor free方式对FCOS检测器做出改进,使得模型拥有anchor based方法的准确率和anchor free方法的推理速度.主要从以下两个方面进行研究:1)以anchor free方法为基础,如何使骨干网络有效提取特征.2)以anchor free方法为基础,在检测器中嵌入特征金字塔网络.从上述两方面,提出了一种基于注意力机制和尺度均衡金字塔网络的目标检测模型.在COCO数据集上,无论是属于anchor based方法的YOLOv3,Faster RCNN,还是属于anchor free方法的Foveabox, FSAF,FCOS,在所提方法的加成下都获得了更高的准确率.所提出的优化模型在ResNet50骨干网络上可以得到39.5%的COCO AP.

In the object detection area, there are two different methods: anchor-based and anchor-free methods, based on whether the prediction box and ground-truth box are associated in advance.Anchor-based method is mostly applied in early object detection, which can improve accuracy, but need complex tricks and increase computation cost, and result in extra model training time and slow inference speed, thus causing significant reduction in application efficiency.The anchor-free method removes the redundant calculation about anchor and improves the inference speed, but the accuracy is slightly sacrificed.In this paper, the advantages of the above methods are combined, based on the anchor-free paradigm we improved the FCOS,which resulted in a model that has the accuracy of the anchor-based method and the inference speed of the anchor-free method.Our study has two main aspects: One is how to extract features effectively in the backbone network based on the anchor-free paradigm, and the other is that feature pyramid network(FPN) is embedded in the detector based on the anchor-free paradigm.From above two aspects, we propose a new object detection method based on attention mechanism and scale equilibrium pyramid network.In COCO dataset, the anchor-based paradigm method including YOLOv3,and Faster RCNN,and the anchor-free paradigm methods including Foveabox, FSAF,and FCOS both achieved higher accuracy within our model.Our optimized model obtains 39.5 % COCO AP on ResNet50 backbone network.