[1]李北明,金荣璐,徐召飞,等.基于特征蒸馏的改进Ghost-Yolov5红外目标检测算法[J].郑州大学学报(工学版),2022,43(01):20-26.[doi:10.13705/j.issn.1671-6833.2022.01.013]
　Li Beiming,Jin Ronglu,Xu Zhaofei,et al.CCFAI:An improved Ghost-YOLOv5 infrared target detection algorithm ba＜x＞sed on feature distillation[J].Journal of Zhengzhou University (Engineering Science),2022,43(01):20-26.[doi:10.13705/j.issn.1671-6833.2022.01.013]
点击复制

基于特征蒸馏的改进Ghost-Yolov5红外目标检测算法()
分享到：

《郑州大学学报(工学版)》[ISSN:1671-6833/CN:41-1339/T]

卷:
43卷
期数:
2022年01期
页码:
20-26
栏目:
出版日期:
2022-01-09

文章信息/Info

Title:
CCFAI:An improved Ghost-YOLOv5 infrared target detection algorithm ba＜x＞sed on feature distillation
作者:
李北明1; 金荣璐1; 2; 徐召飞2; 刘晴2; 王水根2
哈尔滨工程大学信息与通信工程学院;烟台艾睿光电科技有限公司;

Author(s):
Li Beiming; Jin Ronglu; Xu Zhaofei; Liu Qing; Wang Shuigen;
School of Information and Communication Engineering, Harbin University of Engineering; Yantai Arari Optoelectronics Technology Co., Ltd.;

关键词:
Keywords:
分类号:
TP391.9
DOI:
10.13705/j.issn.1671-6833.2022.01.013
文献标志码:
A
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 algorithm has always been one of the import research fields, widely used in security, military and other fields. At present, most of the target detection algorithms suffer from problems such as high computational complexity, low accuracy and poor adaptability, which leads to the problem that the existing approaches are hard to meet pratical applications, especially being deployed on mobile em＜x＞bedded platforms. Towards solving the above problems, in this paper, we proposed an improved Ghost-YOLOv5 infrared target detection algorithm ba＜x＞sed on feature distillation. Firstly, Ghost-net block is adopted for YOLOv5 backbone pruning, and then channel pruning is used to compact the whole network one step further. Secondly, two effective data enhancement methods including Mosaic and Copy-paste, together with feature distillation, are employed to improve the accuracy in ob＜x＞ject detection. Further, that there doesnt exist public dataset in the field of security infrared target detection, this paper constructs an infrared image dataset (coming soon) that contains a variety of scenarios with pedestrians, motor vehicles, and non motorized targets. Experimental results of tests conducted on the above dataset show that compared with original YOLOv5, the proposed algorithm has the following improvements: (1) 93% smaller with the model size as only 1.9MB (2) 86.1% lower complexity with 6.2 GFLOPS and (3) 0.3 accuracy higher with 90.4 overall mAP. We also found that the Mosaic, Copy-paste data enhancement and feature distillation could improve the overall mAP by 6.9. The testing results on Hisi platform with model Hi3519AV100 show that our model achieves higher accuracy and faster running speed with 90.2 mAP and 25 FPS, which is 0.4 higher than YOLOv5s, 4.8 higher than YOLOv3, 14.4 higher than YOLOv3-tiny, 9.6 higher than YOLOv4-tiny, in terms of accuracy.
更新日期/Last Update: 2022-01-09