基于多任务分支SSD的目标检测算法

Object detection algorithm based on multitask branch SSD

针对原始SSD算法各检测特征层没有关联导致特征融合较差,使得检测效果不佳,而现有改进算法DSSD以及RSSD等检测速度太慢的问题,提出一种基于多任务分支的SSD目标检测算法。对特征金字塔进行研究,构建语义与定位级联模块和融合分裂模块用于两个不同分支,在通过两个分支模块之后得到两组多尺度特征,构建多尺度通道聚合模块进行融合和加权,得到最终用于检测的特征金字塔。实验结果表明,在PASCAL VOC 2007数据集上达到79.6%的检测精度,与SSD、DSSD相比具有更好的准确率,检测速度优于DSSD,具有实时检测的能力。

Aiming at the problem that the detection feature layers of the original SSD algorithm are not associated,resulting in poor feature fusion and poor detection effect,and the detection speed of the existing improved algorithms DSSD and RSSD are too low,an object detection algorithm based on multi-task branch SSD was proposed.The feature pyramid was studied,and the semantic and positioning cascade module and the fusion splitting module were constructed for two different branches.After pas-sing through the two branch modules,two sets of multi-scale features were obtained,and a multi-scale channel aggregation mo-dule was constructed for fusing and weighting to get the final feature pyramid for detection.Experimental results show that the detection accuracy of 79.6%is achieved on the PASCAL VOC 2007 data set,which is better than SSD and DSSD,and the detection speed is also higher than DSSD with the ability of real-time detection.