Search

Pruning is recently prevalent in deep neural network compression to save memory footprint and accelerate network inference. Unstructured pruning, i.e., fine-grained pruning, helps preserve model accuracy, while structural pruning, i.e., coarse-grained pruning, is preferred for general-purpose platforms such as GPUs. This paper proposes a regularization-free structural pruning scheme to take advantage of both unstructured and structural pruning by heuristically mixing vector-wise fine-grained and block-wise coarse-grained pruning masks with an AND operation. Experimental results demonstrate that the proposal can achieve higher model accuracy and higher sparsity ratio of VGG-16 on CIFAR-10 and CIFAR-100 compared with commonly applied block and balanced sparsity.

In this paper, we study the problem of interrupted synthetic aperture radar (SAR) imaging and coherent change detection (CCD) in the setting of gapped collections with missing pulses. Conventional interrupted SAR CCD technique is performed using a coherence estimator for paired SAR images which are reconstructed by matched-filtering (MF). However, this method suffers from substantial false alarms due to the poor resolution and high sidelobe level problems in the reconstruction. To improve image resolution and lessen the number of false alarms, sparsity-driven (SD) estimation techniques have been utilized to form SAR image pair and then apply coherence estimator to the resulting images. Actually, existing SD-based regularization SAR imaging algorithms are sensitive to the regularized parameters which are required to tune manually, leading to a serious restriction in practical application. To solve this problem, in this work, we develop an adaptive joint imaging and CCD algorithm in interrupted environments based on Bayesian framework. Our formulation utilizes a partially coherent model to incorporate prior information about the scenes and properties of changes. In particular, we model the sparsity and spatial clustering in changes by virtue of a Markov random fields (MRF) prior. To tackle the difficulty of the calculation of posterior, the mean-field variational Bayesian expectation-maximization (VBEM) method is utilized to simultaneously estimate the MRF parameters and the latent variables. Experimental results are provided to verify the effectiveness of the proposed method.

The performance of synthetic aperture radar (SAR) reconstruction is significantly deteriorated by the random phase noises arising from the atmospheric turbulence or frequency jitter of the transmit signal. Recently, the emerging phase retrieval (PR) technique is gradually extended to the SAR reconstruction problem via the phase-corrupted data attributing to its alluring potential for phase noise mitigation. In this paper, a novel PR-based SAR reconstruction algorithm for phase noise mitigation is proposed by jointing alternating direction method of multipliers (ADMM) and Kolmogorov spectral factorization (KoSF). Owing to the exploiting of the hidden convexity of PR-based SAR reconstruction problem and the structure advantage of the quadratic magnitude measurement, the proposed algorithm acquires better robustness for the complex-valued Gaussian white noises and the random phase noises than the existing PR-based SAR reconstruction algorithms. In the experiments, the synthetic scene data and the moving and stationary target recognition Sandia laboratories implementation of cylinders (MSTAR SLICY) target data are provided to verify the validity of the proposed algorithm.

To identify potential man-made objects, traditional synthetic aperture radar (SAR) imaging techniques are used to consider the dependence in either frequency or azimuth domain. However, these existing methods may result in inaccurate scattering estimation without utilising the inherent polarimetric diversity property. To improve the imaging performance, a polarimetric object-level SAR imaging algorithm is proposed. In the scheme, the polarimetric SAR imaging with canonical scattering characterisation is transferred into a simultaneous sparse approximation (SSA) problem by virtual of incorporating sparse representation and canonical shape feature model. Then the SSA problem is solved via an efficient l 2 , p -norm ( p ∈ ( 0 , 1 ) ) regularisation algorithm. The main advantages of the proposed method are twofold: (i) considering the dispersive, anisotropic and polarimetric scattering characteristics of scatterers allows for more accurate estimation of physically relevant scattering geometry information of scattering centres in comparison with the method ignoring polarisation dependence; and (ii) by exploiting joint sparsity of the multiple polarisation measurements, it can effectively enhance the recovery accuracy in the consistency of the canonical scatterers’ number and locations in different channels. Experimental results are provided to verify the effectiveness of the proposed method.

Random phase noises arising from frequency jitter of transmit signal and atmospheric turbulence result in corrupted synthetic aperture radar (SAR) imagery, which in turn degrades change detection (CD) performance. In this paper, a phase retrieval (PR) based SAR reconstruction autofocus framework by exploiting the hidden convexity is proposed with the goal of achieving reliable persistent surveillance CD. Firstly the original non-convex quartic SAR reconstruction is reformulated as a convex quadratic program. Under the minimum phase assumption, the auto-correlation retrieval- Kolmogorov factorization (CoRK) algorithm is then utilized to optimally and efficiently retrieve the underlying SAR reflectivity. The devised scheme possesses effective capabilities of phase noise mitigation, thus has a superior CD performance. Experimental results are provided to verify the effectiveness of the proposed method.

In this paper, we propose an efficient joint imaging and coherent change detection (CCD) algorithm to cope with the issue of interrupted synthetic aperture radar (SAR) environ­ments. Information about the changing scenes is considered using a partially coherent model, where the structure characteristics of changes are modeled by a Markov random fields (MRF) prior. Then the variational Bayesian expectation-maximization (VBEM) algorithm is employed to simultaneously approximate the posterior distributions of the change map and scene estimates. The proposed scheme has a superior change detection performance over the classical coherent change detectoris. Representative simulations are conducted to demonstrate the validity of the devised algorithm.

Large-scale datasets have successively proven their fundamental importance in several research fields, especially for early progress in some emerging topics. In this paper, we focus on the problem of visual speech recognition, also known as lipreading, which has received increasing interest in recent years. We present a naturally-distributed large-scale benchmark for lip reading in the wild, named LRW-1000, which contains 1,000 classes with 718,018 samples from more than 2,000 individual speakers. Each class corresponds to the syllables of a Mandarin word composed of one or several Chinese characters. To the best of our knowledge, it is currently the largest word-level lipreading dataset and also the only public large-scale Mandarin lipreading dataset. This dataset aims at covering a "natural" variability over different speech modes and imaging conditions to incorporate challenges encountered in practical applications. It has shown a large variation in this benchmark in several aspects, including the number of samples in each class, video resolution, lighting conditions, and speakers' attributes such as pose, age, gender, and make-up. Besides providing a detailed description of the dataset and its collection pipeline, we evaluate several typical popular lipreading methods and perform a thorough analysis of the results from several aspects. The results demonstrate the consistency and challenges of our dataset, which may open up some new promising directions for future work., Comment: IEEE FG 2019

This paper aims at minimizing economical cost of a microgrid by jointly scheduling various devices, e.g., appliances, batteries, thermal generators, and wind turbines. To properly model the system, the characteristics of all devices is fully investigated; in particular, the chance constraint is introduced to capture the randomness of power generation of wind turbines. Then, this problem is formulated as a large-scale mixed-integer program with coupling constraints. In order to solve this problem efficiently, it is decoupled via dual decomposition into a set of subproblems to be solved distributedly on each appliance, battery, and generator. While the scheduling of generator is well studied in literature, this paper specially proposes an efficient method for appliance scheduling, and then employs Benders’ decomposition for battery scheduling. The performance of the proposed approach is verified by numerical simulations.

The recently proposed single image super-resolution method which makes use of the theory of sampling signals with finite rate of innovation (FRI) can obtain the state-of-the-art performance. However, the FRI-based method might bring in artifacts in border areas because some values of edge pixels might be overcorrected, leading to extremely sharp edges. To address such issues, we propose an improved FRI-based single image super-resolution method. Specifically, our method can detect the artifacts, and correct the fake value by taking advantage of the correlation of the surrounding pixels. The method would prevent the edges of image from becoming extremely sharp and preserve the true edge texture. The obtained simulation results reveal that the improved FRI-based method can achieve a better performance than the original FRI-method.

Ground-based synthetic aperture radar (GB-SAR) interferometry is an increasingly popular terrestrial remote sensing technique for landslide monitoring. However, in actual application, the atmospheric phase screen (APS) seriously affects the accuracy of GB-SAR interferometry. Therefore, it is necessary to estimate the APS accurately. This paper develops an improved APS compensation method based on a 2D-polynomial regression model (2D-PRM) with a two-step estimation process using GB-SAR measurements. In this paper, permanent scatterer (PS) points which are selected by a joint method using amplitude, amplitude deviation and correlation coefficient are used to roughly estimate the model parameters. Then, by imposing the condition of the standard deviation of the fitting error, the outliers whose values enormously affect the estimation of parameters are filtered out of the estimation process. A more efficient estimation of the APS is finally achieved by using the remaining pixels based on 2D-PRM. At last, a landslide monitoring experiment is carried out to verify the performance of the proposed method.

The performance of Synthetic aperture radar (SAR) imagery is often significantly deteriorated by the random phase noises arose from the atmospheric turbulence or frequency jitter of the transmit signal within SAR observations. The computational time of the traditional phase retrieval based SAR autofocus algorithms is sharply increased with the size of scene. In this paper, we recast the SAR imaging problem via the phase-corrupted data as a special case of block-based quadratic compressed sensing (BBQCS) problem. We propose a novel fast SAR imaging algorithm to recover the focused well SAR image from the phase-corrupted data and reduce the computational time and memory requirement for several orders of magnitude. Experimental results show our proposed algorithm not only reduces the computational complex but also provides satisfactory reconstruction performance.

Super-resolution (SR) algorithms for multiview videos aim at refining the low-resolution (LR) views using high-frequency detail information from adjacent high-resolution (HR) views. However, to refine a LR view, the state-of-the-art methods only exploit the spatial redundancy of a multiview video. To take full advantage of the spatial and temporal redundancy, a new spatiotemporal SR algorithm is proposed in this paper. Depth-image-based rendering (DIBR) establishes coordinate correspondences between different views. Motion estimation handles object motion between different frames. The high-frequency details are extracted in transform domain and added to LR images. In order to merge the details from different HR images, a combination of global and local weights is proposed. The effectiveness of the proposed algorithm is proved theoretically and shown experimentally.

Foreign Object Debris (FOD) on airport runway is one of the greatest threats to aviation safety. The major characteristics of FOD are diversity and smallness. Synthetic aperture radar (SAR) has a natural advantage in high resolution imaging, and is, therefore, very suitable for small and weak targets imaging on airport runway. Firstly, a new millimeter-wave FOD imaging system called Spotlight Circular Synthetic Aperture Radar (SC-SAR) is introduced in this paper. The SC-SAR system not only can solve the practical limitations for ground-based applications but also fully exploit the advantage of existing ground-based SAR systems for FOD imaging. Secondly, based on the geometry and imaging principle of SC-SAR, the adjusted version of the Range-Doppler imaging formula was presented. Finally, the simulation results validate feasibility of SC-SAR imaging system and algorithm.

This paper focuses on the load shifting problem in a household scenario with a large-capacity battery. We propose a novel Model Predictive Control (MPC) framework to control the charge/discharge power of battery, hence to shave the peak load. Being different from other studies, the framework is designed on the base of individual habit of energy consumption, as it is envisioned that the individual habit is critical for choosing the suitable energy services. In this paper, the habit is modeled as a Markov process and gradually learned by an iterative algorithm; thus, the habit can be utilized for the prediction of future energy consumption. Then, the rolling optimization is applied for the optimal control of the charge/discharge power of battery. It is shown by numerical simulations that the proposed approach can significantly reduce the peak load.

A synchronous Nyquist folding receiver (SNYFR) could intercept wideband signals in multi-Nyquist zones with two analog-to-digital converters. The instantaneous Nyquist zone (INZ) of the SNYFR may not exist or is not unique, and the dual local harmonics was presented to improve the stability of the SNYFR. It was found that the existence and unique conditions of the INZ determined that the frequency of the local oscillator signal (LOS) in one zone was constant. A symmetrical low-pass filter and dual harmonics LOSs were adopted to facilitate the detection of the INZ. Simulations showed that the proposed structure is valid for the interception of wideband radar signal. Key words: Electronic countermeasures, parameter estimation, linear frequency modulation, synchronous Nyquist folding receiver.

An intelligence receiver quasi-matched to the received LPI radar signals in time and frequency domain is presented. The architecture and prototype using FPGA and DSP are shown. The presented receiver works especially well when the received signals are complex and time-varying.

A new algorithm for the recognition of three kinds of intra-pulse modulation signals is presented in this paper, which is based on the theoretical analysis of phase difference and delay correlation dechirp. The local peak values of the signal's instantaneous frequency, computed by phase difference, are derived based on the values of the phase discontinuities. These values can be used to identify the binary phase shift keying (BPSK) signal. The fast recognition between monopluse and linear frequency modulation (LFM) signal can be realized based on the estimation of the chirp rate, thus avoided unwrapping the frequency ambiguity caused by signal-to-noise ratio (SNR) threshold, meanwhile finished the fast parameter estimation of the LFM signal. (4 pages)

In this study, to detect person-borne concealed threats in range profiles under the circumstance of unknown clutter, we propose a binary integration nonparametric detection method based on the generalized sign (GS) detector for range-spread targets in a distributed terahertz radar network (DTRN). In the detection, the length of range-spread targets and the number of dominant scatterers on range-spread targets are considered and adaptively estimated. Furthermore, the GS detection method is applied to maintain a constant false alarm rate (CFAR) under the circumstance of unknown clutter. The detection performance of the proposed method for single terahertz radar and DTRN are both examined with the data synthesized by real range-spread targets data and real clutter data. Experimental results show that the proposed method is effective, and for a given false alarm probability, the DTRN exhibits better detection performance than the single terahertz radar.