The L-band geosynchronous synthetic aperture radar (GEO SAR) is susceptible to the ionospheric irregularities due to its lower carrier frequency, which can be more complicated because the velocity of drifting irregularities is comparable with the scanning velocity of ionospheric penetration points (IPPs) of GEO SAR. In this paper, a drifting ionospheric scintillation simulator (DISS) is proposed for generating L-band GEO SAR data affected by the drifting ionospheric irregularities. The DISS is divided into a static 2D scintillation transfer function generator, a drifting IPP offset indexer and a GEO SAR echo simulator. Based on the DISS, the simulation results indicate that the drifting ionospheric irregularities of the L-band GEO SAR will present different artifacts and more serious decorrelations in L-band GEO SAR images, compared with the L-band low-Earth-orbit (LEO) SAR.
Synchronization is an inevitable technical challenge in distributed synthetic aperture radar (DiSAR). Current work is mainly focused on configuration with the separated transmitter and receiver. In this paper, another configuration with multi-monostatic SARs is proposed to address the issue of frequency synchronization error. Combined with the system characteristic and mechanism of frequency error, the frequency synchronization error model is established first. The analysis shows that frequency synchronization error can lead to phase error, which in turn results in image degradation. Then, to compensate for the phase error, the paper uses an autofocus algorithm combined with the back projection (BP) algorithm. This algorithm adopts the coordinate descent method and iteratively adjusts the phase error estimation until the image reaches its maximum sharpness. Finally, simulations are carried out to validate the effectiveness of the algorithm. The results show that the used method can improve the imaging quality of the multi-monostatic DiSAR system.
Multilingual meta learning has emerged as a promising paradigm for transferring the knowledge from source languages to facilitate the learning of low-resource target languages. Loss functions are a type of meta-knowledge that is crucial to effective training of neural networks, however, the misalignment between the loss functions and learning paradigms of meta learning degrade the network performance. To address this challenge, we propose a new method called Task-based Meta PolyLoss (TMPL) for meta learning. By regarding speech recognition tasks as normal samples and apply PolyLoss to meta loss function, TMPL can be denoted as a linear combination of polynomial functions based on task query loss. We conduct extensive experiments on four low-resource languages from the IARPA BABEL dataset. The results show a significant improvement in performance when applying our TMPL to meta learning.
In this paper, we propose a frequency domain pesudo-oblique projection generalized sidelobe canceller beamformer (FPOGSC) for ultrasound imaging, which aims to improve the resolution and contrast performance. Firstly, the echo signals are divided into multi-segment narrowband time-frequency signals through short-time fourier transform. Then, in the frequency domain, the corresponding data are selected to calculate the pesudo-oblique projection operator to optimize the weight vector. Finally, using the time-frequency conversion method, the output of time-domain beamformer can be obtained. The simulation results show that the full width at half maxima (FWHM) of the proposed method is 79.3% narrower than that of delay-and-sum (DAS). At the same time, the contrast of FPOGSC is 30.3% and 38.0% higher than that of generalized sidelobe canceller (GSC) and DAS respectively.
Service robots have been widely used in many indoor scenes, but their interaction ability based on action recognition has been developed slowly. In this work, we focus on human action recognition for service robots. We find that the existing action recognition datasets are rarely based on robot perspective, and do not focus on human-robot interaction. Therefore, we propose a multi-modal visual dataset named THU-HRIA dataset on the perspective of service robot, including total of eight human daily actions and interactive actions. Based on this dataset, we generated dataset for model training through data division and expansion, then selected several advanced GCN networks and improved their accuracy through transfer learning and finally compared the impact of various factors on recognition performance. According to the experimental results, we design a prototype of action recognition system, which can carry out end-to-end real-time action recognition on a laptop. We propose a time-frames constraint sampling strategy for this system and prove the feasibility of the system through experiments.
Low-light image enhancement (LLIE) is often used to improve the quality of video images in dark light conditions. For LLIE, global image enhancement is performed on the entire image which often leads to a waste of computational resources and ineffectiveness. In this paper, we proposed a radar-camera fusion framework-based approach to enhance the region of interest (ROI). The radar was utilized to locate target areas in video data for local enhancement. Target location was then projected onto the image based on the spatio-temporal calibration and a bounding box centered on the target was got. The corresponding part of the image within the bounding box was the ROI, and only this part of the image was considered for LLIE. An obvious benefit for image local enhancement is the computational efficiency. Experimental results showed that the local image enhancement scheme improved the visual effect while reducing the computation time. One more inspiring benefit is the improvement of object detection performance. It can be seen that the accuracy and the metric of mAP@0.5:0.95 trained by the network based on locally enhanced images were better than the training results using unprocessed images or globally enhanced images.
Internet-of-Things (IoT) devices are often used to transmit physical sensor data over digital wireless channels. Traditional Physical Layer Security (PLS)-based cryptography approaches rely on accurate channel estimation and information exchange for key generation, which irrevocably ties key quality with digital channel estimation quality. Recently, we proposed a new concept called Graph Layer Security (GLS), where digital keys are derived from physical sensor readings. The sensor readings between legitimate users are correlated through a common background infrastructure environment (e.g., a common water distribution network or electric grid). The challenge for GLS has been how to achieve distributed key generation. This paper presents a Federated multi-agent Deep reinforcement learning-assisted Distributed Key generation scheme (FD2K), which fully exploits the common features of physical dynamics to establish secret key between legitimate users. We present for the first time initial experimental results of GLS with federated learning, achieving considerable security performance in terms of key agreement rate (KAR), and key randomness.
In this paper, a new method for high-resolution ISAR imaging under low SNR using parameterized translational motion compensation is proposed. Firstly, translational motion of the target is modeled as formula of the polynomial coefficient vector. Then, entropy minimization corresponding to range profile with compensation term based on coefficients above is taken as objective function. Moreover, the Golden Section Search (GSS) is utilized to search the optimal parameters, which are estimated precisely and efficiently to implement motion compensation. Experimental results from both simulated and real data verify that the proposed method can achieve high-resolution fine focused ISAR image under low SNR with validity and robustness.
Respiratory diseases remain one of the major problems of public health, and early identification of these diseases benefits patient management, disease treatment and contagion control. During COVID-19 pandemic, using cough sound analysis to classify respiratory diseases seems promising. In this study, a deep learning network (VGGish-BiLSTM-attention) is implemented. It employs pre-trained VGGish structure, bidirectional long-short-term-memory (BiLSTM) and attention module, and besides, an output layer is followed for feature fusion and disease classification. Meanwhile, data augmentation, cough sound representation and transfer learning are used for performance boosting. On the COUGHVID database, a binary classification problem (“maybe-covid” versus “no-covid”) is formed, and the proposed network achieves the state-of-the-art result (accuracy 92.41%; precision, 92.59%; recall, 91.97%; F1-score, 92.23%). The ablation study indicates that data augmentation contributes the most with more than 12% increase, and VGGish, BiLSTM and attention module are also important in cough sound based disease analysis. In the future work, more efforts could be made to finely stratify cough sounds and to design advanced models for accurate disease classification and personalized medicine.
Synthetic aperture radar (SAR) tomography (TomoSAR) can separate the layover scatterers and obtain the position and reflectivity of each scatterer to realize the 3-D reconstruction. The compressive-sensing based algorithms have been widely used in TomoSAR inversion. However, this superior performance comes at large computational cost. In this article, we propose a deep learning based TomoSAR inversion approach, which is the learned iterative shrinkage thresholding algorithm (LISTA). Since the network parameters can be trained, the LISTA network can obtain more accurate reconstruction. The pretrained network can also greatly improve the computational efficiency. The simulation experiments show that the LISTA algorithm has super-resolution performance and can effectively separate the layover scatterers.
Multiple signal classification (MUSIC) is one of the common methods for estimating the direction of arrival (DOA). To enhance the accuracy of DOA estimation results, this paper proposes an improved MUSIC algorithm called Taylor-MUSIC. The horizontal linear array is divided into multiple groups by using a fixed interval, resulting in several sets of horizontal arrays. The MUSIC algorithm is then employed to calculate the spatial spectral functions for each group. The obtained spatial spectral functions are further weighted based on the Taylor series. The DOA of the sound source is effectively estimated by searching for the peak value in the spatial spectral functions. The performance of the DOA estimation method based on the Taylor-MUSIC algorithm is analytically investigated, demonstrating its high estimation accuracy even in low signal-to-noise ratios and scenarios with multiple sources. Moreover, the data analysis results of the SWellEx-96 experiment conducted in 1996 indicate that the Taylor-MUSIC algorithm exhibits a more pronounced spatial spectrum and improved angular resolution in the source direction.
In response to the frequent occurrence of disability incidents resulting from improper home fitness practices, a fitness dance scoring system based on human pose recognition has been developed. The system comprises several modules, including the standard video preprocessing module, action information storage module, user action processing module, action scoring module, follow-along guidance module, and UI interaction module. Keyframe sequences are extracted from fitness dance videos, and coordinate information of skeletal points is extracted using OpenPose. It employs various similarity algorithms such as dynamic time warping and frame difference to calculate the similarity between standard dance motions and user motions. Real-time scoring is provided along with guidance to help users improve their movements. The system has been tested and has demonstrated high real-time performance and accuracy on personal computers. It enables comprehensive and detailed evaluation of user performance, providing ordinary users with a scientific and safe home fitness solution, and effectively preventing disability incidents caused by improper exercise.
Due to the time-varying and non-stationary characteristics of micro-motion signals, micro-Doppler spectrograms have been widely used for micro-motion signature analysis, which plays an important role in civil and military fields. However, the background noise in the spectrogram severely restricts subsequent applications such as feature extraction and parameter estimation. In this paper, we propose a novel micro-motion signal enhancement method based on the multi-scale feature pyramid structure and convolutional autoencoder. Compared to the traditional micro-Doppler spectrogram enhancement, the proposed method can restore micro-motion components and suppress background noise even under low signal-to-noise (SNR) conditions by integrating high-level semantic features and low-level detail information. Experimental results demonstrate the effectiveness and robustness of our method.
This paper proposes a multi-user secret key generation system using off-the-shelf ESP32 modules. We first propose a new differential quantization algorithm and contrast the features of differential quantization with uniform quantization and existing equal-probability quantization. Then, we propose a differential compensation algorithm, which can effectively improve the randomness of differential quantization. The improved randomness is higher than that of equal-probability quantization, and the key disagreement rate (KDR) is lower than that of equal-probability quantization. After that, we create an ESP32-based multi-user secret key generation system. We compare the differences of KDR in three movement modes, stable, people-move and trolly-move, and in two scenarios, line-of-sight (LOS) and non-line-of-sight (NLOS). The results show that the KDR with third-order quantization is less than 0.1 for the LOS scene and less than 0.15 for the NLOS scene.
Target feature extraction in synthetic aperture radar images is a critical aspect of target recognition. However, the presence of noise in target characteristics complicates the task of improving the recognition rate. In this study, we propose a feature extraction algorithm centered around attribute scattering centers. Initially, the watershed algorithm is employed to segment and extract the target region in the synthetic aperture radar image, thereby reducing the computational cost. Subsequently, the characteristic parameters of the target scattering centers are obtained using the attribute scattering center model. Finally, the orthogonal matching pursuit algorithm is applied to complete the extraction of scattering centers. Our results demonstrate that this approach significantly improves the efficiency and accuracy of radar image scattering center extraction. Numerous experiments performed on the MSTAR database attest to the efficacy of the proposed synthetic aperture radar image target scattering center feature extraction method, achieving an average recognition accuracy of 99.31% for three types of targets in the MSTAR database.
In recent years, target speech separation has drawn a lot of attention with the development of deep-learning methods. The target speech from the specific direction-of-interest (DOI) can be extracted with auxiliary directional information. However, separating target signals from DOI has not been investigated in detail, and the performance of existing systems can degrade when the direction estimation error occurs. In this paper, a spatial filtering convolutional recurrent network (SF-CRN) is proposed for target speech separation in the direction-of-interest. Inspired by the GCC-PHAT localization method, we construct direction feature (DF) to integrate with multi-channel short-time complex spectra as the input. In addition, we propose a method to train our network to make it more robust to direction estimation error. The experimental results show that our network can achieve significant performance improvement on target speech separation.
Conventional searching methods with only one constant false alarm rate can no longer provide effective detection performance for different remote targets associated with distinct signal-to-noise ratios in radar systems. To overcome this problem, this paper presents a dynamic joint beam and time resource allocation strategy based on variable false alarm rate detection framework for phased array radar for remote target searching. Specifically, a multi-stage searching framework of the phased array radar is first developed, where the false alarm probability for detection increases adaptively with the stage evolving. Then, a constrained allocation model of the beam and time is formulated. Simulation results demonstrate the effectiveness and superiority of the proposed algorithm against some counterparts in terms of detection performance and efficiency of resource utilization.
This paper deals with the problem of mainlobe jamming suppression for phased-array radar system via a joint polarization-space-time processing method resorting to tensor decomposition. Firstly, a dual-polarization receiving array model is established. Then, base on the polarization-space-time characteristics, CANDECOMP/PARAFAC (CP) decomposition is leveraged to recast the received signals as a third-order tensor consisting of three factor matrices. Further, the target echo and the jamming signal can be obtained via minimizing the Frobenius norm of the approximation error between the factor matrices and their iteration forms. Finally, the effectiveness of the proposed method by numerical simulation is verified, showing its capability to resist mainlobe jamming compared with the art competing method.
This paper presents a bit-level image encryption algorithm, which can protect the important bit-planes with respect to the content of plain-image. The plain-image is initially transform to a binary image and decomposed into 8 bitplanes. Then, the similarities among the binary image and each bitplane are measured, and we get an important index array of the 8 bitplanes. Next, the top k important bitplanes are encrypted. Finally, all the bitplanes are combined into the cipher-image. The experimental results of sufficient security analyses demonstrate that proposed algorithm can offer efficient image protection.
