针对人体姿态估计算法的沉重计算成本和庞大网络规模问题,提出面向人体姿态估计的轻量级高效视觉变换器(lightweight efficient vision transformer for human posture estimation,LEViTPose)。引入深度可分离卷积、通道混洗和多尺度卷...针对人体姿态估计算法的沉重计算成本和庞大网络规模问题,提出面向人体姿态估计的轻量级高效视觉变换器(lightweight efficient vision transformer for human posture estimation,LEViTPose)。引入深度可分离卷积、通道混洗和多尺度卷积核并行技术来设计轻量级预处理模块LStem;提出一种级联组空间线性退化注意力(cascaded group spatial linear reduction attention,CGSLRA),采用特征分组划分多个注意头的方式来提升内存存储效率,采用组内特征降维来降低计算冗余;通过引入逐点卷积和分组反卷积来设计轻量级特征还原模块(lightweight feature recovery module,LFRM)。实验结果表明,所提方法相比基线模型,可以在提升网络性能和推理速度的同时降低网络规模和计算开销。在MPII和COCO验证集上与LiteHRNet-30相比,平均准确率分别提高了2.6和3.4个百分点,推理速度提升了1倍。展开更多
The paper develops a multiple matching attenuation method based on extended filtering in the curvelet domain,which combines the traditional Wiener filtering method with the matching attenuation method in curvelet doma...The paper develops a multiple matching attenuation method based on extended filtering in the curvelet domain,which combines the traditional Wiener filtering method with the matching attenuation method in curvelet domain.Firstly,the method uses the predicted multiple data to generate the Hilbert transform records,time derivative records and time derivative records of Hilbert transform.Then,the above records are transformed into the curvelet domain and multiple matching attenuation based on least squares extended filtering is performed.Finally,the attenuation results are transformed back into the time-space domain.Tests on the model data and field data show that the method proposed in the paper effectively suppress the multiples while preserving the primaries well.Furthermore,it has higher accuracy in eliminating multiple reflections,which is more suitable for the multiple attenuation tasks in the areas with complex structures compared to the time-space domain extended filtering method and the conventional curvelet transform method.展开更多
文摘针对人体姿态估计算法的沉重计算成本和庞大网络规模问题,提出面向人体姿态估计的轻量级高效视觉变换器(lightweight efficient vision transformer for human posture estimation,LEViTPose)。引入深度可分离卷积、通道混洗和多尺度卷积核并行技术来设计轻量级预处理模块LStem;提出一种级联组空间线性退化注意力(cascaded group spatial linear reduction attention,CGSLRA),采用特征分组划分多个注意头的方式来提升内存存储效率,采用组内特征降维来降低计算冗余;通过引入逐点卷积和分组反卷积来设计轻量级特征还原模块(lightweight feature recovery module,LFRM)。实验结果表明,所提方法相比基线模型,可以在提升网络性能和推理速度的同时降低网络规模和计算开销。在MPII和COCO验证集上与LiteHRNet-30相比,平均准确率分别提高了2.6和3.4个百分点,推理速度提升了1倍。
基金funded by the Wenhai Program of the ST Fund of Laoshan Laboratory (No.202204803)the National Natural Science Foundation of China (Nos.42074138,42206195)+1 种基金the National Key R&D Program of China (No.2022YFC2803501)the Research Project of the China National Petroleum Corporation (No.2021ZG02)。
文摘The paper develops a multiple matching attenuation method based on extended filtering in the curvelet domain,which combines the traditional Wiener filtering method with the matching attenuation method in curvelet domain.Firstly,the method uses the predicted multiple data to generate the Hilbert transform records,time derivative records and time derivative records of Hilbert transform.Then,the above records are transformed into the curvelet domain and multiple matching attenuation based on least squares extended filtering is performed.Finally,the attenuation results are transformed back into the time-space domain.Tests on the model data and field data show that the method proposed in the paper effectively suppress the multiples while preserving the primaries well.Furthermore,it has higher accuracy in eliminating multiple reflections,which is more suitable for the multiple attenuation tasks in the areas with complex structures compared to the time-space domain extended filtering method and the conventional curvelet transform method.
