基于一阶段模式的目标检测模型的设计与调优

DESIGN AND OPTIMIZATION OF ONE-STAGE-BASED OBJECT DETECTION MODEL

现有的实时目标检测模型在小物体检测上存在精度不高的问题,通过使用新的网络结构以及更加合理简洁的特征使用方式来优化该问题。以密集型连接的形式设计特征提取网络,利用全部浅层特征来辅助对小物体的检测,并保持其能满足实时性要求;改进YOLOv2检测模型中的浅层特征融合方式,通过多种尺寸的池化操作来简化和优化YOLOv2的重组过程。在数据集PASCAL VOC2007+2012上,取得了75.6%的平均准确率(mAP)和52.7帧/s(实时性要求30帧/s)的检测速度,相比YOLOv2提升0.5%mAP。

The existing real-time object detection models have the problem of low accuracy in small object detection,and we optimized it by using a new network structure and more reasonable and concise method with feature maps.The feature extraction network was designed in the form of dense connection.All the shallow features were used to assist the detection of small objects,and it could meet the real-time requirements.The shallow feature fusion method in the YOLOv2 detection model was improved,and the reorganization process of YOLOv2 was simplified and optimized by pooling operation of various sizes.Experiments conducted on PASCAL VOC2007+2012 show that it gets 75.6 mAP and 52.7 Frames/s(30 Frames/s for real time),which is 0.5%higher than YOLOv2.