基于特征蒸馏的改进Ghost-Yolov5红外目标检测算法-《郑州大学学报(工学版)》

文章信息/Info

Title:: An Improved Ghost-YOLOv5 Infrared Target Detection Algorithm Based on Feature Distillation

Author(s):: LI Beiming¹; JIN Ronglu¹; 2; XU Zhaofei²; LIU Qing²; WANG Shuigen²; 1.College of Information and Communication Engineering, Harbin Engineering University, Harbin 150001, China;
2.Yantai IRay Technology Co., Ltd., Yantai 264000, China

Keywords:: infrared target detection; data enhancement; model pruning; feature distillation; Hisi plat-form

摘要:: 红外目标检测算法一直是重要的研究领域之一，广泛应用于安防、军事等领域。当前大部分目标检测算法存在计算复杂度高、准确率低、适应性差等问题，难以满足实际应用需求，以及部署于移动嵌入式平台设备。针对上述问题，本文提出了一种基于特征蒸馏的改进Ghost-YOLOv5红外目标检测算法，该算法首先在YOLOv5模型结构的基础上利用Ghost模块做模型剪枝，然后继续对模型进行通道剪枝；其次使用Mosaic和Copy-paste两种数据增强方法和特征蒸馏方法提高压缩后模型的检测精度。进一步地，考虑到安防红外目标检测领域尚无开源的数据集，本文构建了一个包含多种场景下人、机动车和非机动车目标的数据集（即将开源）。在上述数据集上测试实验结果表明：本文提出的算法利用Ghost模块和通道裁剪可以把原始的YOLOv5模型大小减少到1.9M，相比于原模型大小减少了93%，模型复杂度（GFLOPS）降低了86.1%，并通过知识蒸馏和数据增强的方法，使得裁剪后的模型在红外数据集上的精度提升了6.9，总体mAP值达到了90.4，比YOLOv5模型高0.3。在型号为Hi3519AV100的海思平台上实测，模型的检测速度能达到25帧，平均检测精度能达到90.2，比YOLOv5s高0.4，比YOLOv3高4.8，比YOLOv3-tiny高14.4，比YOLOv4-tiny高9.6。

Abstract:: Infrared target detection algorithms suffered from problems such as poor adaptability and high computational complexity. An improved Ghost-YOLOv5 infrared target detection algorithm was proposed based on feature distillation to solve the above problems. Firstly, GhostNet block was used for backbone pruning. Se-condly, two effective data enhancement methods including Mosaic and Copy-paste were used, together with feature distillation to improve the accuracy in object detection. Furthermore, an infrared image dataset that contained a variety of scenarios with pedestrians, motor vehicles, and non motorized targets was constructed. The test experimental results on the above dataset showed that the model parameters obtained by the algorithm proposed in this paper using GhostNet module were only 1.9M, and the accuracy of the small model on the infrared dataset were improved by 6.6% through feature distillation and data enhancement. And the overall mAP value reached 90.1%. The detection speed of the model could reach 25 frames per second and the average detection accuracy could reach 90.2% when measured empirically on Hisi, all achieving higher detection accuracy compared to a variety of common models portable to this platform.