针对传统红外图像行人姿态识别的问题,在经典LeNet-5模型的基础上,提出一种改进型LeNet-5的网络模型.网络设定输入红外图像尺寸为256×256× 1,选取4层卷积计算增加网络深度,以Leaky ReLu为激活函数并加入dropout层,最后以1×1卷积代替全连接,减小模型参数尺寸,防止过拟合.实验将改进型LeNet-5与经典LeNet-5模型进行比对,结果表明改进型LeNet-5效果最好.与流行的ShuffleNet,NasNet-mobile,EfficientNet-b0和MobileNetV2算法进行对比,实验结果表明,所得测试集的准确率达到97.5％,mean average precision,average recall和F1-score性能指标均优于其他算法.
Human Pose Recognition Method Based on Deep Convolutional Network in Infrared Images
Aiming at the problem of pedestrian pose recognition in traditional infrared images,we pro-posed an improved LeNet-5 network based on the classic LeNet-5 model.The input infrared image size was set as 256 × 256 × 1.Four layers of convolution was selected to deepen the network depth,Leaky ReLu was used as the activation function and adds the Dropout layer was added.Finally,it uses 1×1 convolution in-stead of full connection was used to reduce the model parameter size and prevent overfitting.Compared the im-proved LeNet-5 model with the classic LeNet-5 model,the experimental results show that the improved LeNet-5 model has the best performance.Compared it with popular ShuffleNet,NasNet-mobile,EfficientNet-b0 and MobileNetV2 algorithms,the results show that the proposed network had better mean average preci-sion,average recall,and F1-score.
万方数据
维普
