基于改进AlexNet的红外图像行人姿态识别
Human Pose Recognition in Infrared Images Based on Improved AlexNet
赵丹 1郭姗姗 1计尚冉 1谢雨晴 1方子睿 1单巍1
作者信息
摘要
针对传统红外图像行人姿态识别准确率低下的问题,在经典AlexNet网络的基础上,提出一种改进型AlexNet网络.该网络设定输入红外图像的尺寸为227x227x3,包含5个卷积层、3个池化层、3个全连接层和1个输出层.同时,减小卷积核提取更精细的特征,减少节点数防止过拟合,删除分组和随机失活神经元操作获得更快的提取和计算速度.实验结果表明,与流行的GA-CNN、CNN-SVM、CNN-MLP、CNN-RF算法对比,改进网络的Mean Precision、Mean Recall和Mean F1等性能指标均优于对比算法.
Abstract
For the problem of low accuracy in pedestrian pose recognition of traditional infrared images,an improved AlexNet network is proposed based on the classic AlexNet network.The network sets the size of the input infrared image to 227 x 227 x 3.It includes 5 convolutional layers,3 pooling layers,3 fully connected layers,and 1 output layer.It reduces the number of convolutional kernels to extract finer features,and nodes to prevent overfitting.Meanwhile,it removes grouping and randomly inactivate neurons for faster extraction and computation speed.Experiments show that,compared with popular GA-CNN,CNN-SVM,CNN-MLP,and CNN-RF algorithms,our networks,such as Mean Precision,Mean Recall,and Mean F1 are superior in the performance indicators
关键词
改进型AlexNet/红外图像/姿态识别/深度学习Key words
improved AlexNet/infrared image/pose recognition/deep learning引用本文复制引用
基金项目
安徽省高等学校科学研究重点项目(2022AH050392)
安徽省质量工程项目(2021jyxm1336)
淮北师范大学质量工程项目(2023jxyj020)
安徽省大学生创新创业训练计划(S202310373089)
安徽省大学生创新创业训练计划(S202310373101)
出版年
2024
NETL
NSTL
万方数据