基于特征金字塔网络的目标检测算法

Object Detection Algorithm Based on Feature Pyramid Network

目标检测是计算机视觉研究的一个主要方向,传统目标检测算法为检测不同尺度的图像,往往需要将图片进行缩放搭建图像金字塔,再在图像金字塔上做目标检测。将图像金字塔应用于基于深度学习的目标检测算法主要存在实时性问题。为此提出在网络中搭建特征金字塔,大大提高网络效率,并且提高检测效果。在PASCAL VOC2007目标检测权威数据库上实验结果表明,提出的算法能大大提高目标检测的召回率。

Object detection is a primary mission in computer vision. To deal with different scale objects, classical methods first con- struct image pyramid, and then do object detection on it. Applying image pyramid in deep learning based object detection al- gorithm is inefficient. For this, proposes a feature pyramid network, highly improve the efficiency of the object detection algo- rithm. Examinations on Pascal VOC 2007 dataset show that this algorithm can improve recall efficiently.