一种小尺度目标检测卷积神经网络设计

A Convolutional Neural Network Architecture for Small-Scale Object Detection

航天装备的智能感知技术受距离与探测手段的制约,面临着检测目标尺度小的问题,深度卷积神经网络是目前目标检测的主要技术手段,但小尺度目标在卷积神经网络前向计算过程中,由于多次下采样的网络结构会损失较多的特征信息而不利于目标检测。特征金字塔网络(feature pyramid network,FPN)是一种广泛应用于小尺度目标检测的网络设计,采用主干网络低层特征与高层特征上采样相融合的方式。提出将特征图放大尺寸的网络设计方法,并对网络分离提升主干网络低/高层特征图与小/大尺度目标的匹配度,设计了一种特征漏斗网络(feature funnel networks,FFN)。经实验验证,特征漏斗网络相较于同级别网络在小尺度目标数据集VisDrone中获得了更高的检测精度与召回率而不损失过多的速度。

Due to the constraint of distance and imaging technique,the intellisense technology for aerospace equipment is facing a problem that objects to be detected is small-scale.Deep Convolutional Neural Networks(CNN)is the most popular technique for object detection,but when the small-scale object is calculated in the inference phase,most feature will be lost because of the network architecture with plenty of down sampling layers.Feature Pyramid Network(FPN)has been widespread used in small-scale object detection,which build high-level semantic feature maps at all scales.In this paper,an architecture with feature map scale up is proposed,separating the high-level and low-level feature maps to match the big-scale and small-scale objects,which is called a Feature Funnel Networks(FFN).It is con-firmed by experiment that compared with other network on the VisDrone dataset,the Feature Funnel Networks has higher mAP performance while does not decrease much more the speed of detection.