多尺度卷积特征融合的SSD目标检测算法

SSD Object Detection Algorithm with Multi-Scale Convolution Feature Fusion

提出了一种改进的多尺度卷积特征目标检测方法,用以提高SSD(single shot multibox detector)模型对中目标和小目标的检测精确度。该方法先对SSD模型低层特征层采用区域放大提取的方法以提高对小目标的检测能力,再对高层特征层进行特征提取以改善中目标的检测效果。最后,利用SSD模型中原有的多尺度卷积检测方法,将改进的多层特征检测结果进行融合,并通过参数再训练以获得最终改进的SSD模型。实验结果表明,该方法在MS COCO数据集上对中目标和小目标的测试精确度分别为75.1%和40.5%,相比于原有SSD模型分别提升16.3%和23.1%。

In this paper, it is proposed that the accuracy of small and medium objects detection of SSD(single shot multibox detector) can be improved by incorporating a modified multi-scale convolution feature fusion method. In order to enhance small objects detection, the region magnification extraction on the low-layer feature maps is adopted. Then, features are extracted from the high-layer feature maps to make the detection of medium objects better. These features are finally fused by the multi-scale convolution detection method as in the original SSD architecture. Moreover, the parameters of the present model are obtained via the parameter retraining. On the MS COCO test set, the present model shows that the mAP(mean average precision) of medium and small objects decetion is 75.1% and 40.5% respectively, which is nearly 16.3% and 23.1% higher than the original SSD.