Based on the CNN-LSTM fusion deep neural network,this paper proposes a seismic velocity model building method that can simultaneously estimate the root mean square(RMS)velocity and interval velocity from the common-mi...Based on the CNN-LSTM fusion deep neural network,this paper proposes a seismic velocity model building method that can simultaneously estimate the root mean square(RMS)velocity and interval velocity from the common-midpoint(CMP)gather.In the proposed method,a convolutional neural network(CNN)Encoder and two long short-term memory networks(LSTMs)are used to extract spatial and temporal features from seismic signals,respectively,and a CNN Decoder is used to recover RMS velocity and interval velocity of underground media from various feature vectors.To address the problems of unstable gradients and easily fall into a local minimum in the deep neural network training process,we propose to use Kaiming normal initialization with zero negative slopes of rectifi ed units and to adjust the network learning process by optimizing the mean square error(MSE)loss function with the introduction of a freezing factor.The experiments on testing dataset show that CNN-LSTM fusion deep neural network can predict RMS velocity as well as interval velocity more accurately,and its inversion accuracy is superior to that of single neural network models.The predictions on the complex structures and Marmousi model are consistent with the true velocity variation trends,and the predictions on fi eld data can eff ectively correct the phase axis,improve the lateral continuity of phase axis and quality of stack section,indicating the eff ectiveness and decent generalization capability of the proposed method.展开更多
Action recognition is important for understanding the human behaviors in the video,and the video representation is the basis for action recognition.This paper provides a new video representation based on convolution n...Action recognition is important for understanding the human behaviors in the video,and the video representation is the basis for action recognition.This paper provides a new video representation based on convolution neural networks（CNN）.For capturing human motion information in one CNN,we take both the optical flow maps and gray images as input,and combine multiple convolutional features by max pooling across frames.In another CNN,we input single color frame to capture context information.Finally,we take the top full connected layer vectors as video representation and train the classifiers by linear support vector machine.The experimental results show that the representation which integrates the optical flow maps and gray images obtains more discriminative properties than those which depend on only one element.On the most challenging data sets HMDB51 and UCF101,this video representation obtains competitive performance.展开更多
Action recognition is an important topic in computer vision. Recently, deep learning technologies have been successfully used in lots of applications including video data for sloving recognition problems. However, mos...Action recognition is an important topic in computer vision. Recently, deep learning technologies have been successfully used in lots of applications including video data for sloving recognition problems. However, most existing deep learning based recognition frameworks are not optimized for action in the surveillance videos. In this paper, we propose a novel method to deal with the recognition of different types of actions in outdoor surveillance videos. The proposed method first introduces motion compensation to improve the detection of human target. Then, it uses three different types of deep models with single and sequenced images as inputs for the recognition of different types of actions. Finally, predictions from different models are fused with a linear model. Experimental results show that the proposed method works well on the real surveillance videos.展开更多
Combining both visible and infrared object information, multispectral data is a promising source data for automatic maritime ship recognition. In this paper, in order to take advantage of deep convolutional neural net...Combining both visible and infrared object information, multispectral data is a promising source data for automatic maritime ship recognition. In this paper, in order to take advantage of deep convolutional neural network and multispectral data, we model multispectral ship recognition task into a convolutional feature fusion problem, and propose a feature fusion architecture called Hybrid Fusion. We fine-tune the VGG-16 model pre-trained on ImageNet through three channels single spectral image and four channels multispectral images, and use existing regularization techniques to avoid over-fitting problem. Hybrid Fusion as well as the other three feature fusion architectures is investigated. Each fusion architecture consists of visible image and infrared image feature extraction branches, in which the pre-trained and fine-tuned VGG-16 models are taken as feature extractor. In each fusion architecture, image features of two branches are firstly extracted from the same layer or different layers of VGG-16 model. Subsequently, the features extracted from the two branches are flattened and concatenated to produce a multispectral feature vector, which is finally fed into a classifier to achieve ship recognition task. Furthermore, based on these fusion architectures, we also evaluate recognition performance of a feature vector normalization method and three combinations of feature extractors. Experimental results on the visible and infrared ship (VAIS) dataset show that the best Hybrid Fusion achieves 89.6% mean per-class recognition accuracy on daytime paired images and 64.9% on nighttime infrared images, and outperforms the state-of-the-art method by 1.4% and 3.9%, respectively.展开更多
针对点云配准过程中,下采样时容易丢失关键点、影响配准精度的问题,本文提出一种基于特征融合和网络采样的配准方法,提高了配准的精度和速度。在PointNet分类网络基础上,引入小型注意力机制,设计一种基于深度学习网络的关键点提取方法,...针对点云配准过程中,下采样时容易丢失关键点、影响配准精度的问题,本文提出一种基于特征融合和网络采样的配准方法,提高了配准的精度和速度。在PointNet分类网络基础上,引入小型注意力机制,设计一种基于深度学习网络的关键点提取方法,将局部特征和全局特征融合,得到混合特征的特征矩阵。通过深度学习实现对应矩阵求解中相关参数的自动优化,最后利用加权奇异值分解(singular value decomposition,SVD)得到变换矩阵,完成配准。在ModelNet40数据集上的实验表明,和最远点采样相比,所提算法耗时减少45.36%;而配准结果和基于特征学习的鲁棒点匹配(robust point matching using learned features,RPM-Net)相比,平移矩阵均方误差降低5.67%,旋转矩阵均方误差降低13.1%。在自制点云数据上的实验,证实了算法在真实物体上配准的有效性。展开更多
基金financially supported by the Key Project of National Natural Science Foundation of China (No. 41930431)the Project of National Natural Science Foundation of China (Nos. 41904121, 41804133, and 41974116)Joint Guidance Project of Natural Science Foundation of Heilongjiang Province (No. LH2020D006)
文摘Based on the CNN-LSTM fusion deep neural network,this paper proposes a seismic velocity model building method that can simultaneously estimate the root mean square(RMS)velocity and interval velocity from the common-midpoint(CMP)gather.In the proposed method,a convolutional neural network(CNN)Encoder and two long short-term memory networks(LSTMs)are used to extract spatial and temporal features from seismic signals,respectively,and a CNN Decoder is used to recover RMS velocity and interval velocity of underground media from various feature vectors.To address the problems of unstable gradients and easily fall into a local minimum in the deep neural network training process,we propose to use Kaiming normal initialization with zero negative slopes of rectifi ed units and to adjust the network learning process by optimizing the mean square error(MSE)loss function with the introduction of a freezing factor.The experiments on testing dataset show that CNN-LSTM fusion deep neural network can predict RMS velocity as well as interval velocity more accurately,and its inversion accuracy is superior to that of single neural network models.The predictions on the complex structures and Marmousi model are consistent with the true velocity variation trends,and the predictions on fi eld data can eff ectively correct the phase axis,improve the lateral continuity of phase axis and quality of stack section,indicating the eff ectiveness and decent generalization capability of the proposed method.
基金Supported by the National High Technology Research and Development Program of China(863 Program,2015AA016306)National Nature Science Foundation of China(61231015)+2 种基金Internet of Things Development Funding Project of Ministry of Industry in 2013(25)Technology Research Program of Ministry of Public Security(2016JSYJA12)the Nature Science Foundation of Hubei Province(2014CFB712)
文摘Action recognition is important for understanding the human behaviors in the video,and the video representation is the basis for action recognition.This paper provides a new video representation based on convolution neural networks（CNN）.For capturing human motion information in one CNN,we take both the optical flow maps and gray images as input,and combine multiple convolutional features by max pooling across frames.In another CNN,we input single color frame to capture context information.Finally,we take the top full connected layer vectors as video representation and train the classifiers by linear support vector machine.The experimental results show that the representation which integrates the optical flow maps and gray images obtains more discriminative properties than those which depend on only one element.On the most challenging data sets HMDB51 and UCF101,this video representation obtains competitive performance.
文摘Action recognition is an important topic in computer vision. Recently, deep learning technologies have been successfully used in lots of applications including video data for sloving recognition problems. However, most existing deep learning based recognition frameworks are not optimized for action in the surveillance videos. In this paper, we propose a novel method to deal with the recognition of different types of actions in outdoor surveillance videos. The proposed method first introduces motion compensation to improve the detection of human target. Then, it uses three different types of deep models with single and sequenced images as inputs for the recognition of different types of actions. Finally, predictions from different models are fused with a linear model. Experimental results show that the proposed method works well on the real surveillance videos.
文摘Combining both visible and infrared object information, multispectral data is a promising source data for automatic maritime ship recognition. In this paper, in order to take advantage of deep convolutional neural network and multispectral data, we model multispectral ship recognition task into a convolutional feature fusion problem, and propose a feature fusion architecture called Hybrid Fusion. We fine-tune the VGG-16 model pre-trained on ImageNet through three channels single spectral image and four channels multispectral images, and use existing regularization techniques to avoid over-fitting problem. Hybrid Fusion as well as the other three feature fusion architectures is investigated. Each fusion architecture consists of visible image and infrared image feature extraction branches, in which the pre-trained and fine-tuned VGG-16 models are taken as feature extractor. In each fusion architecture, image features of two branches are firstly extracted from the same layer or different layers of VGG-16 model. Subsequently, the features extracted from the two branches are flattened and concatenated to produce a multispectral feature vector, which is finally fed into a classifier to achieve ship recognition task. Furthermore, based on these fusion architectures, we also evaluate recognition performance of a feature vector normalization method and three combinations of feature extractors. Experimental results on the visible and infrared ship (VAIS) dataset show that the best Hybrid Fusion achieves 89.6% mean per-class recognition accuracy on daytime paired images and 64.9% on nighttime infrared images, and outperforms the state-of-the-art method by 1.4% and 3.9%, respectively.
文摘针对点云配准过程中,下采样时容易丢失关键点、影响配准精度的问题,本文提出一种基于特征融合和网络采样的配准方法,提高了配准的精度和速度。在PointNet分类网络基础上,引入小型注意力机制,设计一种基于深度学习网络的关键点提取方法,将局部特征和全局特征融合,得到混合特征的特征矩阵。通过深度学习实现对应矩阵求解中相关参数的自动优化,最后利用加权奇异值分解(singular value decomposition,SVD)得到变换矩阵,完成配准。在ModelNet40数据集上的实验表明,和最远点采样相比,所提算法耗时减少45.36%;而配准结果和基于特征学习的鲁棒点匹配(robust point matching using learned features,RPM-Net)相比,平移矩阵均方误差降低5.67%,旋转矩阵均方误差降低13.1%。在自制点云数据上的实验,证实了算法在真实物体上配准的有效性。
