Low-altitude infrared small target detection based on fully convolutional regression network and graph matching

Abstract

The high-precision infrared small target detection under low-altitude background has a high application value. The existing small IR target detection methods usually fail or cause a high probability of false alarm in the highly moving and complex low-altitude backgrounds. A detection method based on a fully convolutional network in spatial and graph matching in temporal is proposed. First, a deep fully convolution regression network for high-precision small target detection is designed to obtain accurate target probability heat maps. A weighted bipartite graph matching model is established for the target trajectory association based on the targets' temporal characteristics. It uses the motion and radiation similarity between the detection results in adjacent frames to eliminate false alarms caused by random noise and clutter. Finally, it further integrates the target optical flow information into the trajectory to distinguish between real targets and fixed ground objects. Many experimental results show that the method in this paper can accurately detect small targets in complex moving backgrounds and achieve a high detection rate in the case of a low false alarm rate.