融合上下文信息特征的小目标检测方法

Small Object Detection Method with Multi-level Feature Fusion of Context Information

由于小目标有限的分辨率和表观信息,其检测任务一直是计算机视觉领域的挑战性工作。在解决这一问题时,现有大多数方法为了提高精度而牺牲了速度。在论文中,为了提高小目标检测精度,同时保证检测速度,提出了一种在卷积网络中引入上下文信息的特征融合方法,即Contextual Fused Network(简称CF-Net)。CF-Net引入了上下文信息,并且只在浅层进行特征融合,这样既能提高小目标的检测精度,又能保证检测速度。实验结果表明,在小目标检测上,CF-Net在PASCAL VOC2007上获得的mAP为78.9,比目前主流的单点检测器SSD提高了2%。CF-Net模型测试速度为40 fps,比现有小目标检测器DSSD高26.4 fps。

Due to its limited resolution and apparent information,small targets have always been challenging tasks in the field of computer vision.To solve this problem,most existing methods sacrifice speed in order to improve accuracy.In this paper,in order to improve the accuracy of detecting small targets while ensuring the detection speed,a feature fusion method named contextual fused network(CF-Net)is proposed.CF-Net introduces context information,and only performs feature fusion at the shallow layer,which can not only improve the detection accuracy of small targets but also ensure the detection speed.The experimental results show that the mAP obtained by CF-Net on PASCAL VOC2007 is 78.9,especially in the detection of small targets,which is 2%higher in accuracy than the current mainstream single shot multi-box detectors.The CF-Net model test speed is 40 fps,which is 26.4 fps higher than the existing small target detector DSSD.