基于AS-YOLOv7的红外目标检测算法
Research on Infrared Target Detection Algorithm Based on AS-YOLOv7
王艳梅 1李明明 1黄月 2张艳珠2
作者信息
- 1. 沈阳理工大学信息科学与工程学院,沈阳 110159
- 2. 沈阳理工大学自动化与电气工程学院,沈阳 110159
- 折叠
摘要
为解决传统目标检测算法在红外场景下的目标识别精度低、检测速度差等问题,提出基于AS-YOLOv7的红外目标检测方法.首先针对红外图像的特性,对YOLOv7模型中的特征提取网络进行重构并构建卷积块CBH,保证模型的稳定性和计算速度;然后改进金字塔池化(SPPCSPC)模块,使模型在不增加任何参数的同时提高模型性能;最后将原模型中的损失函数替换为具有动态非单调的WIoU损失,解决红外数据集中质量较好和质量较差的样本间的不平衡问题,提高模型的泛化性.本文实验在自行采集的数据集上进行训练和验证,与原模型相比,AS-YOLOv7的平均精度提升了 0.2％,参数量降低了 79％,计算量降低了 85％,推理时间降低了32 ms.改进后的模型在保持准确度的情况下提高了模型的推理速度.
Abstract
In order to solve the problems of low target recognition accuracy and poor detection speed of traditional target detection algorithms in infrared scenes,an infrared target detection meth-od based on AS-YOLOv7 is proposed.First,according to the characteristics of infrared images,the feature extraction network in YOLOv7 model is reconstructed and the CBH module is established to ensure the stability and calculation speed of the model.Then,the pyramid pooling(SPPCSPC)mod-ule is improved forbetter model performance without adding any parameters.Finally,the loss func-tion in the original model is replaced by a dynamic non-monotonic WIoU loss to solve the imbal-ance between samples with good quality and poor quality in the infrared data set and improve the generalization of the model.The experiment is validated on the self-collected data set.Compared with the original model,the average accuracy of AS-YOLOv7 is improved by 0.2％,the number of parameters is reduced by 79％,the amount of calculation is reduced by 85％,and the inference time is reduced by 32 ms.The improved model has a higher inference speed while maintaining accuracy.
关键词
红外图像/YOLOv7/目标检测/特征提取Key words
infrared image/YOLOv7/target detection/feature-base引用本文复制引用
基金项目
辽宁省教育厅高等学校基本科研项目(LJKZ0245)
装备预研国防科技重点实验室基金项目()
出版年
2024
NETL
NSTL
万方数据