为深入研究光学遥感图像中的船舶检测问题,提升检测精度和降低模型复杂度,设计基于改进旋转区域卷积和神经网络(Rotational Region Convolutional Neural Networks),R^(2)CNN的两阶段旋转框检测模型。在模型的第一阶段使用水平框作为候...为深入研究光学遥感图像中的船舶检测问题,提升检测精度和降低模型复杂度,设计基于改进旋转区域卷积和神经网络(Rotational Region Convolutional Neural Networks),R^(2)CNN的两阶段旋转框检测模型。在模型的第一阶段使用水平框作为候选区域;在模型第二阶段引入水平框预测分支,并且设计一种间接预测角度的回归模型;在测试阶段进行旋转框非极大值抑制时,设计基于掩码矩阵的旋转框IoU(Intersection over Union)算法。试验结果显示:改进R^(2)CNN模型在HRSC2016(High Resolution Ship Collection 2016)数据集上取得81.0%的平均精确度,相比其他模型均有不同程度的提升,说明改进R^(2)CNN在简化模型的同时能有效提升使用旋转框检测船舶的性能。展开更多
<span style="font-family:Verdana;">Convolutional neural networks, which have achieved outstanding performance in image recognition, have been extensively applied to action recognition. The mainstream a...<span style="font-family:Verdana;">Convolutional neural networks, which have achieved outstanding performance in image recognition, have been extensively applied to action recognition. The mainstream approaches to video understanding can be categorized into two-dimensional and three-dimensional convolutional neural networks. Although three-dimensional convolutional filters can learn the temporal correlation between different frames by extracting the features of multiple frames simultaneously, it results in an explosive number of parameters and calculation cost. Methods based on two-dimensional convolutional neural networks use fewer parameters;they often incorporate optical flow to compensate for their inability to learn temporal relationships. However, calculating the corresponding optical flow results in additional calculation cost;further, it necessitates the use of another model to learn the features of optical flow. We proposed an action recognition framework based on the two-dimensional convolutional neural network;therefore, it was necessary to resolve the lack of temporal relationships. To expand the temporal receptive field, we proposed a multi-scale temporal shift module, which was then combined with a temporal feature difference extraction module to extract the difference between the features of different frames. Finally, the model was compressed to make it more compact. We evaluated our method on two major action recognition benchmarks: the HMDB51 and UCF-101 datasets. Before compression, the proposed method achieved an accuracy of 72.83% on the HMDB51 dataset and 96.25% on the UCF-101 dataset. Following compression, the accuracy was still impressive, at 95.57% and 72.19% on each dataset. The final model was more compact than most related works.</span>展开更多
Modulation recognition is a core technology for non-cooperative communications and spectrum monitoring.It is evident that traditional feature-based machine learning methods lack robustness in low-tomedium Signal-to-No...Modulation recognition is a core technology for non-cooperative communications and spectrum monitoring.It is evident that traditional feature-based machine learning methods lack robustness in low-tomedium Signal-to-Noise Ratio(SNR)environments.Conversely,deep learning models,despite their capacity to enhance recognition accuracy,encounter trade-offs between computational efficiency and recognition precision.The present study focuses on lightweight Convolutional Neural Networks(CNN2)and residual neural networks(ResNet),undertaking a systematic comparison of their modulation recognition performance across the full SNR range to determine model suitability for diverse scenarios.The RML2016.10a dataset,which contains 24 modulation types,was used to conduct experiments with nine SNR gradients ranging from-8 to 24 dB.These experiments were conducted under consistent training parameters in a Mac CPU environment.The results of the study indicate the following:In terms of recognition accuracy,ResNet demonstrates superior performance across the entire SNR range,achieving up to 59.33%higher accuracy than CNN2 in the low-to-medium SNR range(-8 to 8 dB),particularly excelling in complex modulated signal recognition.With regard to computational efficiency,CNN2 demonstrated a significant advantage.This research provides a basis for scenario-specific model selection:Where high recognition accuracy is paramount at low-to-medium SNR,ResNet should be prioritized.In contrast,for edge device deployments with limited computational resources,CNN2 is the optimal choice.展开更多
The upgradation in Deep learning has enabled us to use the advantage of classification techniques into hyperspectral image for analysis of variety of image bands.In recent works,convolutional neural network(CNN)are pr...The upgradation in Deep learning has enabled us to use the advantage of classification techniques into hyperspectral image for analysis of variety of image bands.In recent works,convolutional neural network(CNN)are predominantly used for visual processing of data and its classification.Hyperspectral image classification deals with both spatial and spectral information extracted from the image datasets.3-D CNN is sparsely used due to their complexity in computation.However,it has the capability to facilitate both spatial and spectral data from the spectrum of remotely sensed images.Usage of 2-D CNN includes only the spatial information and cannot be feasible for classification.The proposed model includes 3-D CNN stacked with 2-D CNN reducing the complexity.This hybrid model validated for the remote sensing datasets of Indian Pines(IP),Pavia University(PU),Salinas(SA)and Kennedy Space Canter(KSC).The conventional principal component analysis(PCA)is applied to remove the spectral redundancy over the HSI datasets.The model incorporates both 2-D and 3-D CNN,collectively called the Hybrid Spectral CNN.It comprises of three 3-D convolutional layers and one 2-D convolution layer with 3FC(fully connected)layers.The results of the Hybrid model are compared with the 2-D CNN and 3-D CNN models.Several statistical tests like F1-score,Recall,Precision etc.,are also computed.All these experiments show that the performance of hyperspectral image classification is improved efficiently with this framework.展开更多
文摘<span style="font-family:Verdana;">Convolutional neural networks, which have achieved outstanding performance in image recognition, have been extensively applied to action recognition. The mainstream approaches to video understanding can be categorized into two-dimensional and three-dimensional convolutional neural networks. Although three-dimensional convolutional filters can learn the temporal correlation between different frames by extracting the features of multiple frames simultaneously, it results in an explosive number of parameters and calculation cost. Methods based on two-dimensional convolutional neural networks use fewer parameters;they often incorporate optical flow to compensate for their inability to learn temporal relationships. However, calculating the corresponding optical flow results in additional calculation cost;further, it necessitates the use of another model to learn the features of optical flow. We proposed an action recognition framework based on the two-dimensional convolutional neural network;therefore, it was necessary to resolve the lack of temporal relationships. To expand the temporal receptive field, we proposed a multi-scale temporal shift module, which was then combined with a temporal feature difference extraction module to extract the difference between the features of different frames. Finally, the model was compressed to make it more compact. We evaluated our method on two major action recognition benchmarks: the HMDB51 and UCF-101 datasets. Before compression, the proposed method achieved an accuracy of 72.83% on the HMDB51 dataset and 96.25% on the UCF-101 dataset. Following compression, the accuracy was still impressive, at 95.57% and 72.19% on each dataset. The final model was more compact than most related works.</span>
文摘Modulation recognition is a core technology for non-cooperative communications and spectrum monitoring.It is evident that traditional feature-based machine learning methods lack robustness in low-tomedium Signal-to-Noise Ratio(SNR)environments.Conversely,deep learning models,despite their capacity to enhance recognition accuracy,encounter trade-offs between computational efficiency and recognition precision.The present study focuses on lightweight Convolutional Neural Networks(CNN2)and residual neural networks(ResNet),undertaking a systematic comparison of their modulation recognition performance across the full SNR range to determine model suitability for diverse scenarios.The RML2016.10a dataset,which contains 24 modulation types,was used to conduct experiments with nine SNR gradients ranging from-8 to 24 dB.These experiments were conducted under consistent training parameters in a Mac CPU environment.The results of the study indicate the following:In terms of recognition accuracy,ResNet demonstrates superior performance across the entire SNR range,achieving up to 59.33%higher accuracy than CNN2 in the low-to-medium SNR range(-8 to 8 dB),particularly excelling in complex modulated signal recognition.With regard to computational efficiency,CNN2 demonstrated a significant advantage.This research provides a basis for scenario-specific model selection:Where high recognition accuracy is paramount at low-to-medium SNR,ResNet should be prioritized.In contrast,for edge device deployments with limited computational resources,CNN2 is the optimal choice.
文摘The upgradation in Deep learning has enabled us to use the advantage of classification techniques into hyperspectral image for analysis of variety of image bands.In recent works,convolutional neural network(CNN)are predominantly used for visual processing of data and its classification.Hyperspectral image classification deals with both spatial and spectral information extracted from the image datasets.3-D CNN is sparsely used due to their complexity in computation.However,it has the capability to facilitate both spatial and spectral data from the spectrum of remotely sensed images.Usage of 2-D CNN includes only the spatial information and cannot be feasible for classification.The proposed model includes 3-D CNN stacked with 2-D CNN reducing the complexity.This hybrid model validated for the remote sensing datasets of Indian Pines(IP),Pavia University(PU),Salinas(SA)and Kennedy Space Canter(KSC).The conventional principal component analysis(PCA)is applied to remove the spectral redundancy over the HSI datasets.The model incorporates both 2-D and 3-D CNN,collectively called the Hybrid Spectral CNN.It comprises of three 3-D convolutional layers and one 2-D convolution layer with 3FC(fully connected)layers.The results of the Hybrid model are compared with the 2-D CNN and 3-D CNN models.Several statistical tests like F1-score,Recall,Precision etc.,are also computed.All these experiments show that the performance of hyperspectral image classification is improved efficiently with this framework.
