一种基于共享多头模块的轻量型一阶段网络

A lightweight one-stage network based on shared multi-head modules

一阶段目标检测网络SSD备受青睐,本文基于标准的SSD网络,提出了一种新颖的轻量型SSD(Lightweight SSD,LSSD)网络构架;此外,还提出了SMHM(Shared Multi-Head Module)模块,该模块使得所有输出层级共享网络头部。相对于标准的SSD,本文提出的这种改进型一阶段网络构架(记为SMHM-LSSD)具有更少的网络参数量、更快的速度、更高的检测精度。本文在香港中文大学和商汤科技推出的平台mmdetection上对VOC0712数据集进行实验,其中VOC0712训练集进行训练,VOC07测试集进行测试。实验结果显示,本文提出的LSSD相比于SSD提高了0.2%的检测精度,减少了23.1 M的参数量,提升了5 fps/s的速度;加入SMHM模块后,最高提升了0.6%的检测性能,减少28.7 M的参数量,提升8 fps/s的速度,SMHM-LSSD达到了78.9%的均值平均精度。

The one-stage object detection network SSD is very popular. Based on the standard SSD network, this paper proposes a novel LSSD(Lightweight SSD) network architecture. In addition, the SMHM(Shared Multi-Head Module) is proposed, which enables all output levels to share the network head. Compared with the standard SSD, the improved one-stage network architecture(denoted as SMHM-LSSD) proposed in this paper has fewer network parameters, faster speed, and higher detection accuracy. Experiments are conducted on the VOC0712 dataset on the platform mmdetection launched by the Chinese University of Hong Kong and SenseTime. The VOC0712 training set is used for training and the VOC07 test set is used for testing. The experimental results show that the LSSD proposed in this paper improves the detection accuracy by 0.2% compared with the SSD, reduces the parameter amount of 23.1 M, and improves the speed of 5 fps/s. After adding the SMHM module, the detection performance is improved by up to 0.6%. By reducing the parameter amount of 28.7 M and increasing the speed of 8 fps/s, SMHM-LSSD reached 78.9% of mAP.