[1]张 瑶,潘志松*.GP-YOLOX:无预训练的轻量级红外目标检测模型[J].计算机技术与发展,2022,32(12):165-172.[doi:10. 3969 / j. issn. 1673-629X. 2022. 12. 025]
　ZHANG Yao,PAN Zhi-song*.GP-YOLOX:Light-weight Infrared Object Detection Model without Pre-training[J].,2022,32(12):165-172.[doi:10. 3969 / j. issn. 1673-629X. 2022. 12. 025]
点击复制

GP-YOLOX:无预训练的轻量级红外目标检测模型()
分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
32
期数:
2022年12期
页码:
165-172
栏目:
人工智能
出版日期:
2022-12-10

文章信息/Info

Title:
GP-YOLOX:Light-weight Infrared Object Detection Model without Pre-training
文章编号:
1673-629X(2022)12-0165-08
作者:
张 瑶; 潘志松*
陆军工程大学 指挥控制工程学院,江苏 南京 210006
Author(s):
ZHANG Yao; PAN Zhi-song*
School of Command & Control Engineering,Army Engineering University of PLA,Nanjing 210006,China
关键词:
模型轻量化; YOLOX; Ghost 模块; 无预训练剪枝; 目标检测
Keywords:
light-weight model; YOLOX; Ghost module; pruning without pre-training; object detection
分类号:
TP183
DOI:
10. 3969 / j. issn. 1673-629X. 2022. 12. 025
摘要:
YOLOX 是 YOLO 系列最新目标检测算法,不仅实现了超越 YOLOV3、YOLOv4 和 YOLOv5 的检测精度,而且取得了极具竞争力的端到端推理速度。 然而 YOLOX 在嵌入式设备上部署时仍存在模型体积大、浮点数运算量高、实时性不佳等问题,为了解决以上问题,同时避免模型预训练带来的不必要能耗,提出了一种无需预训练的 GP-YOLOX 算法。 该算法首先利用轻量级的 Ghost 模块重构 YOLOX 目标检测网络,初步压缩模型体积,减少运算量;随后对重构后的网络进行无预训练剪枝,选择合适的稀疏比例,在保留精度的前提下,最大化压缩模型体积,减少模型计算量,缩短模型的前向推理时间。首先在 FLIR ADAS 和 KAIST 红外数据集上,对 YOLOX 四种规模的模型进行了实验,最终在保持原有精度的前提下,参数量和浮点数运算量均减小了约 75% ,同时前向推理时间缩短了约 60% ;随后用轻量级网络 MobileNetv3 代替 YOLOX 的骨干网络 DarkNet,与 GP-YOLOX 进行了对比,该方法在同等数量级的参数量和计算量下,明显优于 MobileNetv3。
Abstract:
YOLOX is the latest object detection algorithm of YOLO series,which not only performs better than YOLOv3,YOLOv4 andYOLOv5,but also achieves competitive end to end reasoning speed. However,when YOLOX is deployed on embedded devices,there arestill many problems such as huge parameter size of model,high computation of floating point calculation and poor real-time performance.In order to solve the problems mentioned above and avoid the unnecessary energy consumption in pre-training process,we propose a GP-YOLOX algorithm without pre - training. The algorithm reconstructed YOLOX by lightweight Ghost module, initially compressed themodel volume and reduced the computation. Then,the reconstructed network was pruned without pre-training. Appropriate sparse ratiowas selected to maximize the compression model volume,reduce the model calculation amount and shorten the forward reasoning time ofthe model under premise of maintaining the accuracy. In this paper,reconstructed experiments are conducted on four scale models ofYOLOX on the FLIR ADAS and KAIST infrared datasets. On the premise of maintaining the original accuracy,the amount of parameterand floating-point operations were reduced by about 75% ,while the inference time was reduced by about 60% . Then,MobileNetv3 wasused to replace DarkNet, the backbone of YOLOX. Compared with GP - YOLOX, the proposed method is obviously better than　MobileNet　v3 under the same order of parameters and calculation.

相似文献/References:

[1]刘立峰,李尔康,李 梦,等.基于轻量化模型的工程量统计方法研究[J].计算机技术与发展,2021,31(增刊):202.[doi:10. 3969 / j. issn. 1673-629X. 2021. S. 041]
　LIU Li-feng,LI Er-kang,LI Meng,et al.Research on Statistical Method of Engineering Quantity Based on Lightweight Model[J].,2021,31(12):202.[doi:10. 3969 / j. issn. 1673-629X. 2021. S. 041]
[2]刘立峰,李尔康,孙摇 阳,等.基于直线三维模型剖切方法的研究与实践[J].计算机技术与发展,2021,31(增刊):155.[doi:10. 3969 / j. issn. 1673-629X. 2021. S. 031]
　LIU Li-feng,LI Er-kang,SUN Yang,et al.Research and Practice of Cutting Method Based on Straight Line 3D Model[J].,2021,31(12):155.[doi:10. 3969 / j. issn. 1673-629X. 2021. S. 031]

更新日期/Last Update: 2022-12-10