提出一种基于深度残差网络的轻量级指静脉识别算法。首先,以ResNet34为基础,使用深度可分离卷积代替传统卷积,加入SE(Squeeze and Excitation)注意力机制模块来提取手指静脉空间域上的细节特征,并引入宽度缩放因子,进一步压缩网络;其次...提出一种基于深度残差网络的轻量级指静脉识别算法。首先,以ResNet34为基础,使用深度可分离卷积代替传统卷积,加入SE(Squeeze and Excitation)注意力机制模块来提取手指静脉空间域上的细节特征,并引入宽度缩放因子,进一步压缩网络;其次,在训练中引入教师-学生网络模式,对轻量级深度残差网络进行知识蒸馏训练,并使用知识蒸馏损失、CurricularFace和交叉熵损失对网络进行联合监督,解决了轻量级深度残差网络因学习参数量较少引起的性能下降问题。分别在FV-USM数据集、Lab-Normal数据集和Lab-Special数据集上进行仿真实验,结果表明,同基于轻量级网络MobileFaceNet的识别算法相比,提出的算法有效提高了零误识识别率和Top1排序性能。展开更多
This paper is devoted to temperature analysis on power RF LDMOS with different feature parameters of die thickness, pitch S length and finger width. The significance of these three parameters is determined from temper...This paper is devoted to temperature analysis on power RF LDMOS with different feature parameters of die thickness, pitch S length and finger width. The significance of these three parameters is determined from temperature comparison obtained by 3D Silvaco-Atlas device simulator. The first three simulations focus on temperature variation with the three factors at different output power density respectively. The results indicate that both the thinner die thickness and the broaden pitch S length have distinct advantages over the shorter finger width. The device, at the same time, exhibits higher temperature at a larger output power density. Simulations are further carried out on structure with combination of different pitch s length and die thickness at a large 1W/mm output power density and the temperature reduction reaches as high as 55%.展开更多
文摘提出一种基于深度残差网络的轻量级指静脉识别算法。首先,以ResNet34为基础,使用深度可分离卷积代替传统卷积,加入SE(Squeeze and Excitation)注意力机制模块来提取手指静脉空间域上的细节特征,并引入宽度缩放因子,进一步压缩网络;其次,在训练中引入教师-学生网络模式,对轻量级深度残差网络进行知识蒸馏训练,并使用知识蒸馏损失、CurricularFace和交叉熵损失对网络进行联合监督,解决了轻量级深度残差网络因学习参数量较少引起的性能下降问题。分别在FV-USM数据集、Lab-Normal数据集和Lab-Special数据集上进行仿真实验,结果表明,同基于轻量级网络MobileFaceNet的识别算法相比,提出的算法有效提高了零误识识别率和Top1排序性能。
文摘This paper is devoted to temperature analysis on power RF LDMOS with different feature parameters of die thickness, pitch S length and finger width. The significance of these three parameters is determined from temperature comparison obtained by 3D Silvaco-Atlas device simulator. The first three simulations focus on temperature variation with the three factors at different output power density respectively. The results indicate that both the thinner die thickness and the broaden pitch S length have distinct advantages over the shorter finger width. The device, at the same time, exhibits higher temperature at a larger output power density. Simulations are further carried out on structure with combination of different pitch s length and die thickness at a large 1W/mm output power density and the temperature reduction reaches as high as 55%.
