Objective To propose a Light-Atten-Pose-based algorithm for classifying abnormal morphology in traditional Chinese medicine(TCM)inspection to solve the problem of relying on manual labor or expensive equipment with pe...Objective To propose a Light-Atten-Pose-based algorithm for classifying abnormal morphology in traditional Chinese medicine(TCM)inspection to solve the problem of relying on manual labor or expensive equipment with personal subjectivity or high cost.Methods First,this paper establishes a dataset of abnormal morphology for Chinese medicine diagnosis,with images from public resources and labeled with category labels by several Chinese medicine experts,including three categories:normal,shoulder abnormality,and leg abnormality.Second,the key points of human body are extracted by Light-Atten-Pose algo-rithm.Light-Atten-Pose algorithm uses lightweight EfficientNet network and polarized self-attention(PSA)mechanism on the basis of AlphaPose,which reduces the computation amount by using EfficientNet network,and the data is finely processed by using PSA mecha-nism in spatial and channel dimensions.Finally,according to the theory of TCM inspection,the abnormal morphology standard based on the joint angle difference is defined,and the classification of abnormal morphology of Chinese medical diagnosis is realized by calculat-ing the angle between key points.Accuracy,frames per second(FPS),model size,parameter set(Params),and giga floating-point operations per second(GFLOPs)are chosen as the eval-uation indexes for lightweighting.Results Validation of the Light-Atten-Pose algorithm on the dataset showed a classification accuracy of 96.23%,which is close to the original AlphaPose model.However,the FPS of the improved model reaches 41.6 fps from 16.5 fps,the model size is reduced from 155.11 MB to 33.67 MB,the Params decreases from 40.5 M to 8.6 M,and the GFLOPs reduces from 11.93 to 2.10.Conclusion The Light-Atten-Pose algorithm achieves lightweight while maintaining high ro-bustness,resulting in lower complexity and resource consumption and higher classification accuracy,and the experiments prove that the Light-Atten-Pose algorithm has a better overall performance and has practical application in the pose estimation task.展开更多
In tensor theory, the parallel factorization （PARAFAC）decomposition expresses a tensor as the sum of a set of rank-1tensors. By carrying out this numerical decomposition, mixedsources can be separated or unknown sys...In tensor theory, the parallel factorization （PARAFAC）decomposition expresses a tensor as the sum of a set of rank-1tensors. By carrying out this numerical decomposition, mixedsources can be separated or unknown system parameters can beidentified, which is the so-called blind source separation or blindidentification. In this paper we propose a numerical PARAFACdecomposition algorithm. Compared to traditional algorithms, wespeed up the decomposition in several aspects, i.e., search di-rection by extrapolation, suboptimal step size by Gauss-Newtonapproximation, and linear search by n steps. The algorithm is ap-plied to polarization sensitive array parameter estimation to showits usefulness. Simulations verify the correctness and performanceof the proposed numerical techniques.展开更多
针对小目标水漂垃圾形态多变、分辨率低且信息有限,导致检测效果不理想的问题,提出一种改进的Faster-RCNN(Faster Regions with Convolutional Neural Network)水漂垃圾检测算法MP-Faster-RCNN(Faster-RCNN with Multi-scale feature an...针对小目标水漂垃圾形态多变、分辨率低且信息有限,导致检测效果不理想的问题,提出一种改进的Faster-RCNN(Faster Regions with Convolutional Neural Network)水漂垃圾检测算法MP-Faster-RCNN(Faster-RCNN with Multi-scale feature and Polarized self-attention)。首先,建立黄河兰州段小目标水漂垃圾数据集,将空洞卷积结合ResNet-50代替原来的VGG-16(Visual Geometry Group 16)作为主干特征提取网络,扩大感受野以提取更多小目标特征;其次,在区域生成网络(RPN)利用多尺度特征,设置3×3和1×1的两层卷积,补偿单一滑动窗口造成的特征丢失;最后,在RPN前加入极化自注意力,进一步利用多尺度和通道特征提取更细粒度的多尺度空间信息和通道间依赖关系,生成具有全局特征的特征图,实现更精确的目标框定位。实验结果表明,MP-Faster-RCNN能有效提高水漂垃圾检测精度,与原始Faster-RCNN相比,平均精度均值(mAP)提高了6.37个百分点,模型大小从521 MB降到了108 MB,且在同一训练批次下收敛更快。展开更多
近年来,随着深度学习的发展,在自然街景下的文本检测取得了巨大的进步,但在多方向和弯曲文本及对比度低的文本检测中的效果仍不理想。因此,针对弯曲文本和对比度低的文本的检测问题,提出了一种融合多尺度模块的文本检测方法,并通过检测...近年来,随着深度学习的发展,在自然街景下的文本检测取得了巨大的进步,但在多方向和弯曲文本及对比度低的文本检测中的效果仍不理想。因此,针对弯曲文本和对比度低的文本的检测问题,提出了一种融合多尺度模块的文本检测方法,并通过检测效果的提升,提高端到端文本识别的识别效果。针对RFB(Receptive Field Block)模块在下采样后局部信息丢失的问题,在RFB模块中嵌入极化自注意力(Polarized Self-Attention)机制以改进RFB来提取有效文本特征,提高特征图表征效果。针对特征金字塔(FPN)提取的特征不足、感受野小的问题,将改进的RFB模块嵌入特征金字塔(FPN)模块以增强特征提取融合。针对特征分布不确定性及远距离特征融合效果不佳的问题,引入条形池化(Strip Pooling)模块,进而提升检测方法的鲁棒性。在公开数据集Total-Text上的实验结果表明,该算法的F-measure值在端到端文本识别没有词汇表的情形下与目前高效的MaskTextSpotterV3相比高了0.3百分点,而在有词汇表的情形下则高出了0.2百分点;而在仅文本检测的情形下,该方法也有较为良好的表现。展开更多
基金National Key Research and Development Program of China(2022YFC3502302)。
文摘Objective To propose a Light-Atten-Pose-based algorithm for classifying abnormal morphology in traditional Chinese medicine(TCM)inspection to solve the problem of relying on manual labor or expensive equipment with personal subjectivity or high cost.Methods First,this paper establishes a dataset of abnormal morphology for Chinese medicine diagnosis,with images from public resources and labeled with category labels by several Chinese medicine experts,including three categories:normal,shoulder abnormality,and leg abnormality.Second,the key points of human body are extracted by Light-Atten-Pose algo-rithm.Light-Atten-Pose algorithm uses lightweight EfficientNet network and polarized self-attention(PSA)mechanism on the basis of AlphaPose,which reduces the computation amount by using EfficientNet network,and the data is finely processed by using PSA mecha-nism in spatial and channel dimensions.Finally,according to the theory of TCM inspection,the abnormal morphology standard based on the joint angle difference is defined,and the classification of abnormal morphology of Chinese medical diagnosis is realized by calculat-ing the angle between key points.Accuracy,frames per second(FPS),model size,parameter set(Params),and giga floating-point operations per second(GFLOPs)are chosen as the eval-uation indexes for lightweighting.Results Validation of the Light-Atten-Pose algorithm on the dataset showed a classification accuracy of 96.23%,which is close to the original AlphaPose model.However,the FPS of the improved model reaches 41.6 fps from 16.5 fps,the model size is reduced from 155.11 MB to 33.67 MB,the Params decreases from 40.5 M to 8.6 M,and the GFLOPs reduces from 11.93 to 2.10.Conclusion The Light-Atten-Pose algorithm achieves lightweight while maintaining high ro-bustness,resulting in lower complexity and resource consumption and higher classification accuracy,and the experiments prove that the Light-Atten-Pose algorithm has a better overall performance and has practical application in the pose estimation task.
基金supported by the National Natural Science Foundation of China(61571131)the Technology Innovation Fund of the 10th Research Institute of China Electronics Technology Group Corporation(H17038.1)
文摘In tensor theory, the parallel factorization （PARAFAC）decomposition expresses a tensor as the sum of a set of rank-1tensors. By carrying out this numerical decomposition, mixedsources can be separated or unknown system parameters can beidentified, which is the so-called blind source separation or blindidentification. In this paper we propose a numerical PARAFACdecomposition algorithm. Compared to traditional algorithms, wespeed up the decomposition in several aspects, i.e., search di-rection by extrapolation, suboptimal step size by Gauss-Newtonapproximation, and linear search by n steps. The algorithm is ap-plied to polarization sensitive array parameter estimation to showits usefulness. Simulations verify the correctness and performanceof the proposed numerical techniques.
文摘针对小目标水漂垃圾形态多变、分辨率低且信息有限,导致检测效果不理想的问题,提出一种改进的Faster-RCNN(Faster Regions with Convolutional Neural Network)水漂垃圾检测算法MP-Faster-RCNN(Faster-RCNN with Multi-scale feature and Polarized self-attention)。首先,建立黄河兰州段小目标水漂垃圾数据集,将空洞卷积结合ResNet-50代替原来的VGG-16(Visual Geometry Group 16)作为主干特征提取网络,扩大感受野以提取更多小目标特征;其次,在区域生成网络(RPN)利用多尺度特征,设置3×3和1×1的两层卷积,补偿单一滑动窗口造成的特征丢失;最后,在RPN前加入极化自注意力,进一步利用多尺度和通道特征提取更细粒度的多尺度空间信息和通道间依赖关系,生成具有全局特征的特征图,实现更精确的目标框定位。实验结果表明,MP-Faster-RCNN能有效提高水漂垃圾检测精度,与原始Faster-RCNN相比,平均精度均值(mAP)提高了6.37个百分点,模型大小从521 MB降到了108 MB,且在同一训练批次下收敛更快。
文摘近年来,随着深度学习的发展,在自然街景下的文本检测取得了巨大的进步,但在多方向和弯曲文本及对比度低的文本检测中的效果仍不理想。因此,针对弯曲文本和对比度低的文本的检测问题,提出了一种融合多尺度模块的文本检测方法,并通过检测效果的提升,提高端到端文本识别的识别效果。针对RFB(Receptive Field Block)模块在下采样后局部信息丢失的问题,在RFB模块中嵌入极化自注意力(Polarized Self-Attention)机制以改进RFB来提取有效文本特征,提高特征图表征效果。针对特征金字塔(FPN)提取的特征不足、感受野小的问题,将改进的RFB模块嵌入特征金字塔(FPN)模块以增强特征提取融合。针对特征分布不确定性及远距离特征融合效果不佳的问题,引入条形池化(Strip Pooling)模块,进而提升检测方法的鲁棒性。在公开数据集Total-Text上的实验结果表明,该算法的F-measure值在端到端文本识别没有词汇表的情形下与目前高效的MaskTextSpotterV3相比高了0.3百分点,而在有词汇表的情形下则高出了0.2百分点;而在仅文本检测的情形下,该方法也有较为良好的表现。
