基于一阶段目标检测网络头部算法研究

Research on head algorithm of network based on one-stage object detection

目标检测的网络框架对目标检测结果影响极大,其中网络头部的研究是网络框架改进的重点之一。本文针对一阶段目标检测的网络头部进行改进。通过对当前两阶段网络头部的研究与一阶段网络框架RetinaNet头部热力图的输出进行分析,在一阶段网络头部创新性地引入池化层模块、提出双分类头模块、使用2个网络头部权重自适应分配结合的方法。本文使用RetinaNet作为baseline、VOC0712和MS COCO2017数据集作为实验数据集,最终在VOC0712上mAP达到了80.8%,相比于baseline提高了3.5%,在MS COCO2017测试集上mAP达到了40.2%,相比于RetinaNet提高了1.1%,使用多尺度后mAP达到了41.7%,提高了2.4%。

The network framework of object detection has a great influence on the object detection results,and the research of the network head is one of the focuses of the improvement of the network framework.This paper improves the network header of one-stage object detection.By analyzing the current two-stage network head research and the output of the one-stage network framework RetinaNet head heatmap,the paper innovatively introduces the pooling layer module in the first-stage network head,proposes a dual-classification head module,and uses two networks.After that,a combined method is used for adaptive allocation of head weights.This paper uses RetinaNet as the baseline,VOC0712 and MS COCO2017 datasets as experimental datasets,thereafter mAP achieves 80.8%on VOC0712,which is 3.5%higher than baseline.Meanwhile mAP reaches 40.2%on MS COCO2017 test set,and compared with RetinaNet,it is improved by 1.1%.Furtherly after using multi-scale,mAP reaches 41.7%,which is an increase of 2.4%.