为提高复杂背景下红外小目标的检测能力,提出了一种改进SSD(single shot multiBox detector)与DSST(discriminative scale space tracker)的红外小目标检测算法。在SSD基础网络中加入通道-空间注意力机制与特征金字塔FPN(feature pyramid network)算法,增强深层网络的语义信息,优化目标感受野,利用卷积计算目标特征信息来增强小目标的检测能力;通过尺度判别的DSST方法,解决目标丢失问题,实现连续帧目标的稳定检测。针对小目标在复杂背景下的检测,在PC端和嵌入式ZYNQ平台搭建了能够进行目标融合、目标轨迹设定、目标识别与跟踪的模拟系统,通过不同目标在复杂背景下的叠加融合实现预设场景的高效模拟与测试,避免复杂外场测试实验,评测算法效能。实验分析表明,所提算法在检测存在其他移动目标和干扰噪点的背景区域时,能够有效识别小目标,不同背景区域内多帧检测的AP(average precision)值为98.17%,相较于传统的SSD算法提升了11.28%,体现了算法的有效性。

To improve the detection ability of infrared small targets in complex background, an enhanced algorithm based on SSD and DSST is proposed. The channel space attention mechanism and FPN algorithm are added to the SSD basic network to enhance the semantic information of the deep network, optimize the target receptive field, and use convolution to calculate the target feature information to enhance the detection ability of small targets. Depending on DSST method of scale discrimination, the problem of target loss is solved, and the stable detection of continuous frames is realized. A simulation system which comprises target fusion, target trajectory setting, target recognition and tracking is built. It is combining PC with embedded ZYNQ platform for the detection of small targets under complex background. The preset scenes using for efficient simulation and testing are constructed based on the superposition and fusion of different targets under complex background, which avoids complex field verification experiments. The efficiency for the algorithm is evaluated in simulation system. The experimental validating test shows that the proposed algorithm could effectively identify small targets when detecting background areas with other moving targets and interference noise. The AP value of multi frame detection in different background areas is 98.17%, which is 11.28% higher than the traditional SSD algorithm, reflecting the effectiveness of the algorithm.