融合注意力机制的轻量化大熊猫目标检测模型
A lightweight giant panda object detection model with attention mechanism
吕皓天 1贾小林1
作者信息
- 1. 西南科技大学计算机科学与技术学院,绵阳市移动物联射频识别技术重点实验室,四川绵阳 621010
- 折叠
摘要
针对大熊猫在复杂环境中难以识别,以及目标检测模型在计算资源有限的嵌入式设备上部署效果差的问题,提出一种轻量化的大熊猫目标检测模型Giant Panda-YOLOv5n(GP-YOLOv5n).该模型在YOLOv5n的基础上进行改进,提出了融合注意力的深度可分离颈部网络,提高复杂环境下对目标的检测精度与检测速度,并在边界框回归损失函数中引入Alpha-IoU来提高模型对目标的边界框定位精度.在自制大熊猫数据集上进行训练后,对模型进行嵌入式设备优化并迁移部署到Jetson Nano.实验结果表明,改进后的模型在mAP50和mAP50:95指标上达到了 97.8％和73.6％,相较于原始模型分别提升了 2.7％和9.2％,在嵌入式设备上的模型检测速度达到了 15.12f/s,能够准确、实时地检测出复杂环境中的大熊猫个体.
Abstract
To address the problem of poor detection performance of giant pandas in complex environments and the low efficiency of object detection models on resource-limited embedded devices,a lightweight giant panda detection model called GP-YOLOv5n is proposed.The model is based on YOLOv5n and improved by introducing a depth-sepa-rable neck network with attention,which enhances the detection accuracy and speed of targets in complex environ-ments.Moreover,the model adopts Alpha-IoU in the bounding box regression loss function to improve the bounding box localization accuracy of targets.After training on a homemade giant panda dataset,the model is optimized for em-bedded devices and deployed on Jetson Nano.Experimental results show that the improved model achieves 97.8％and 73.6％in mAP50 and mAP50:95 metrics respectively,which are 2.7％and 9.2％higher than the original model.The detection speed of the model on embedded devices reaches 15.12 f/s,which can accurately and real-time detect the giant pandas in complex environments.
关键词
目标检测/图像处理/嵌入式系统/轻量化Key words
object detection/image processing/embedded system/lightweight引用本文复制引用
基金项目
国家自然科学基金面上项目(61471306)
四川省自然科学基金面上项目(2022NSFSC0548)
四川省重点研发计划项目(2020YFS0360)
四川省教育厅人才培养质量和教学改革项目(JG2021-1414)
出版年
2024
NETL
NSTL
万方数据