Abstract: Deep learning has been attracting increasing attention in the recent decade throughout science and engineering due to its wide range of successful applications. In real problems, however, most implementation stages for applying deep learning still require inevitable manual interventions, which naturally conducts difficulty in its availability to general users with less expertise and also deviates from the intelligence of humans. It is thus a challenging while critical issue to enhance the level of automation across all elements of the entire deep learning framework, like input amelioration, model designing and learning, and output adjustment. This paper tries to list several representative issues of this research topic, and briefly describe their recent research progress and some related works proposed along this research line. Some specific challenging problems have also been presented.
Abstract: Face liveness detection, as a key module of real face recognition systems, is to distinguish a fake face from a real one. In this paper, we propose an improved Convolutional neural network (CNN) architecture with two bypass connections to simultaneously utilize low-level detailed information and high-level semantic information. Considering the importance of the texture information for describing face images, texture features are also adopted under the conventional recognition framework of Support vector machine (SVM). The improved CNN and the texture feature based SVM are fused. Context information which is usually neglected by existing methods is well utilized in this paper. Two widely used datasets are used to test the proposed method. Extensive experiments show that our method outperforms the state-of-the-art methods.
Abstract: Arm motion control is fundamental for robot accomplishing complicated manipulation tasks. Different movements can be organized by configuring a series of motion units. Our work aims at equipping the robot with the ability to carry out Basic unit movements (BUMs), which are used to constitute various motion sequences so that the robot can drive its hand to a desired position. With the definition of BUMs, we explore a learning approach for the robot to develop such an ability by leveraging deep learning technique. In order to generate the BUM regarding to the current arm state, an internal inverse model is developed. We propose to use Conditional generative adversarial networks (CGANs) to establish the inverse model to generate the BUMs. The experimental results on a humanoid robot PKU-HR6.0II illustrate that CGANs could successfully generate multiple solutions given a BUM, and these BUMs can be used to constitute further reaching movement effectively.
Abstract: To solve the red jujube classification problem, this paper designs a convolutional neural network model with low computational cost and high classification accuracy. The architecture of the model is inspired by the multi-visual mechanism of the organism and DenseNet. To further improve our model, we add the attention mechanism of SE-Net. We also construct a dataset which contains 23,735 red jujube images captured by a jujube grading system. According to the appearance of the jujube and the characteristics of the grading system, the dataset is divided into four classes:invalid, rotten, wizened and normal. The numerical experiments show that the classification accuracy of our model reaches to 91.89%, which is comparable to DenseNet-121, InceptionV3, InceptionV4, and Inception-ResNet v2. Our model has real-time performance.
Abstract: Great efforts have been made by using deep neural networks to recognize multi-label images. Since multi-label image classification is very complicated, many studies seek to use the attention mechanism as a kind of guidance. Conventional attention-based methods always analyzed images directly and aggressively, which is difficult to well understand complicated scenes. We propose a global/local attention method that can recognize a multi-label image from coarse to fine by mimicking how human-beings observe images. Our global/local attention method first concentrates on the whole image, and then focuses on its local specific objects. We also propose a joint max-margin objective function, which enforces that the minimum score of positive labels should be larger than the maximum score of negative labels horizontally and vertically. This function further improve our multi-label image classification method. We evaluate the effectiveness of our method on two popular multi-label image datasets (i.e., Pascal VOC and MS-COCO). Our experimental results show that our method outperforms state-of-the-art methods.
Abstract: Cyclic codes of dimension 2 over a finite field are shown to have at most two nonzero weights. We compute their weight distribution, and give a condition on the roots of their check polynomials for them to be maximum distance separable code.
Abstract: Sufficient and necessary conditions for Hermitian (1+λu)-constacyclic self-orthogonal codes over Fpm+uFpm are obtained, where λ is a unit of Fpm+uFpm. Based on this, a new method for the construction of pm-ary quantum codes from the obtained Hermitian (1+λu)-constacyclic self-orthogonal codes over Fpm+uFpm is given. As an example,[[pm-1, pm-2d+1, d]]pm quantum MDS codes are constructed for 1 ≤ d ≤ (pm+1)/2 and p≠2, as well as some other quantum codes.
Abstract: Signal complexity denotes the intricate patterns hidden in the complicated dynamics merging from nonlinear system concerned. The chaotic signal complexity measuring in principle combines both the information entropy of the data under test and the geometry feature embedded. Starting from the information source of Shannon's entropy, combined with understanding the merits and demerits of 0-1 test for chaos, we propose new compression entropy criteria for identifying chaotic signal complexity in periodic, quasi-periodic or chaotic state, in mapping results in 3s-graph with significant different shape of good or bad spring and in Construction creep (CC) rate with distinguishable value-range of[0, 7%], (7%, 50%] or (50%, 84%]. The employed simulation cases are Lorenz, Li and He equations' evolutions, under key information extracting rules of both two-layer compression functions and self-similarity calcu-lation, compared with methods of 0-1 test for chaos, Lyapunov exponent and Spectral Entropy complexity. The research value of this work will provide deep thinking of the concise featureexpressions of chaotic signal complexity measure in feature domain.
Abstract: Fixed polarity Reed-Muller (RM) expression (FPRM) has several practical applications due to its multitude of properties. In order to generate an FPRM with minimum power, based on a genetic algorithm, we propose a Power optimization approach (POA-FPRMs) of Fixed Polarity RM expressions for incompletely specified Boolean functions. Simulation results on MCNC benchmark circuits show that POAFPRMs can effectively reduce power, compared with the traditional polarity optimization approach, where the don't care terms are neglected.
Abstract: To achieve causality reasoning of aviation safety events based on big data of cross-media network, a data-driven general diagnostic framework based on nonaxiomatic logic is designed and implemented. On the basis of this framework, the uncertain causality between aviation safety events and faults is expressed in the form of binary non-axiomatic incident experience at first. A general expression for calculating the attribution and confidence degrees in the non-axiomatic incident experience is given based on records of aviation safety historical incident. A concept of non-axiomatic incident experience graph is proposed, a diagnosis algorithm for aviation safety events is given with the combination of revision and deduction rules in non-axiomatic logic. Experimental results of a Version 1.0 beta demo show that this framework can effectively diagnose all potential faults according to aviation safety events; compared with other machine learning frameworks, it has higher reliability (especially scalability) under the premise of ensuring diagnosis accuracy.
Abstract: Range reduction is the initial and essential stage of function computation, but its pipelined implementation has the drawbacks of large cost and terrible accuracy. We proposed low cost and accurate pipelined range reduction, which adopts truncated multiplier with optimized bit-width to reduce the cost of pipelined implementation and achieve the accuracy within 1 unit in the last place (ulp). TCORDIC algorithm is a widely used algorithm to compute floating-point sine/cosine function, and we implemented the combination of TCORDIC and our range reduction algorithms, verifying the goal of accuracy within 1 ulp.
Abstract: Compared with the method by Zhang in 2017, an extended one for constructing Boolean functions with multiple-valued Walsh spectra is given, which is derived from a bent function by complementing the values at some points. Based on which, new classes of Boolean functions with four-valued and five-valued Walsh spectra are presented, and some of their cryptographic properties are studied.
Abstract: With the popularity of mobile terminals, Quick response (QR) code, which acts as a convenient and quick way to transfer graphic information, has been widely used in various fields. QR code is a 2D matrix code which contains information in both vertical and horizontal directions. In order to promote the characters encoding efficiency, we propose the improved character encoding methods based on QR code. Specifically, we divide the types of encoding data into Chinese characters and nonChinese characters, and thus optimizing their encoding methods separately:we apply the branch-bound algorithm to optimize the encoding pattern selection of QR code; as for Chinese characters, based on the Chinese characters use frequency generated by Url crawler system, we design a new encoding method for Chinese characters. Our approach has been shown to afford better experimental test results compared with traditional encoding methods of QR code.
Abstract: As one of the most commonly used features, Mel-frequency cepstral coefficients (MFCCs) are less discriminative at high frequency. A novel technique, known as Deep scattering spectrum (DSS), addresses this issue and looks to preserve greater details. DSS feature has shown promise both on classification and recognition tasks. In this paper, we extend the use of DSS feature for acoustic scene classification task. Results on Detection and classification of acoustic scenes and events (DCASE) 2016 and 2017 show that DSS provided 4.8% and 17.4% relative improvements in accuracy over MFCC features, within a state-of-the-art time delay neural network framework.
Abstract: Most of the existed vein features are lack of robustness to light intensity variation, and some algorithms rely on the specified vein data sets, which leads to the limitation of real applications. To solve the problems, we propose a novel vein recognition algorithm based on Nonnegative matrix factorization (NMF) with double regularization terms. The innovations of our algorithm are mainly reflected in the following two aspects:in order to improve feature robustness, a novel feature mapping function is designed to map the initial Histogram of oriented gradient (HOG) feature to a new space; to enhance the recognition performance, an effective NMF model is presented, which not only reduces feature dimension, but also optimizes the feature sparsity and clustering property simultaneously. Experiments show that the proposed algorithm can achieve satisfactory results in terms of False rejection rate (FRR) and False acceptance rate (FAR), which indicates that our algorithm is valuable for other classification problems.
Abstract: Inspired by the recent advances in generative networks, we propose a VAE-GAN based hashing framework for fast image retrieval. The method combines a Variational autoencoder (VAE) with a Generative adversarial network (GAN) to generate content preserving images for pairwise hashing learning. By accepting real image and systhesized image in a pairwise form, a semantic perserving feature mapping model is learned under a adversarial generative process. Each image feature vector in the pairwise is converted to a hash codes, which are used in a pairwise ranking loss that aims to preserve relative similarities on images. Extensive experiments on several benchmark datasets demonstrate that the proposed method shows substantial improvement over the state-of-the-art hashing methods.
Abstract: This paper introduces an integrated approach to estimate the fetal heart rate from the abdominal Electrocardiogram (ECG) signal. Empirical mode decomposition (EMD) can decompose the fetal ECG signal into a set of intrinsic mode functions, which can be used as the indicator of the occurrence for the fetal heartbeats. The decomposition basis functions are directly derived from the fetal signal under test, which make the detection process robust and adaptive. Multiple signal classification (MUSIC) is a high resolution algorithm for frequency estimation, which can be applied to the fetal heartbeats indicator sequence output from the preceding EMD, estimating the fetal heart rate in frequency domain without the heartbeat wave detection. Compared with the popular Independent component analysis (ICA) method, the proposed method has shown improved robustness and fidelity in estimation of the fetal heart rate during testing with real fetal ECG database from DaISy and PhysioNet.
Abstract: Hashing for nearest neighbor search has attracted considerable interest recently given its efficiency in speed and storage. Many methods follow a projection-quantization framework which firstly projects original data into low-dimensional compact space and secondly quantifies each projected dimension to 1 bit by thresholding. The variance of projected dimensions, however, may vary a lot so that quantifying them equivalently degrades the searching performance. In this paper, we put forward a novel method, dubbed Balanced hashing (BH), which finds adjustment functions to reproject the data such that the variance of dimensions can be balanced by directly and explicitly maximizing the degree of balance of data, while preserving important properties. Experiments on benchmarks demonstrate that BH can outperform several state-of-the-art methods.
Abstract: Nonlinear feedback shift registers (NFSRs) are widely used in communication and cryptography. How to construct more NFSRs with maximal periods, which can generate sequences with maximal periods, i.e., de Brujin sequences, is an attractive problem. Recently many results on constructing de Bruijn sequences from adjacency graphs of Linear feedback shift registers (LFSRs) by means of the cycle joining method have been obtained. In this paper we discuss a class of LFSRs with characteristic polynomial p2(x), where p(x) is a primitive polynomial of degree n ≥ 2 over the finite field F2. As results, we determine their cycle structures and adjacency graphs, and further construct a class of new de Bruijn sequences from these LFSRs.
Abstract: Focusing on the problem of state estimation in the presence of sensor faults and Out-of-sequence measurement (OOSM) observations synchronously, we derive a formulation of Linear minimum mean squared error (LMMSE) filter with the arbitrary time delay of OOSM, the generalization of the present work lies in simultaneous treatment of correlated faults together with OOSMs that arrive at an arbitrary delay in a linear-optimal manner. The approach is demonstrated in a numerical comparison.
Abstract: The proposed Clock and data recovery system (CDRS) has three improved parts. The second order digital filter with rounding algorithm implements fractional gain and avoids direct current quantization noise which varies between -q/2 and +q/2 while that of traditional filter varies between 0 and +q (q is quantization step). The hysteresis majority voter can combat high frequency and strong jitter especially in quasi-steady state. The improved Phase interpolator (PI) has much smaller current-switching glitch and phase glitch since the weighting current changes gradually instead of steeply. The optimized CDRS can handle up to±6000ppm (parts per million) frequency offset and the phase resolution is 1.4o/LSB (Least significant bit) according to analysis. The simulations of jitter transfer function and jitter tolerance by Matlab, simulations of phase noise by spectre using Verilog+VeriloA model, and measurements of frequency offset and jitter tolerance all show its good performance.
Abstract: Real-time network resource allocation based on virtualization technology is an important method to solve the solidification problem of Fiber-wireless (FiWi) access networks. To increase the resource utilization and meet the specific QoS requirements of smart gird communication services, a Load-balancing and QoS based dynamic resource allocation method (LbQ-DR) is proposed with three sub-mechanisms. A time-window based substrate network resource update mechanism is designed to describe the realtime resource consumption of substrate networks, which can balance the accuracy of resource status and the complexity of the allocation algorithm. A QoS-based Virtual network request (VNR) sorting mechanism is presented to precisely calculate the priority of VNRs and reasonably sort the incoming services. A load-balancing based resource allocation mechanism is designed to avoid unbalanced resource consumption. Especially, the channel interference is considered in the cost of embedding and a collision domain mechanism is introduced to decrease interference. Simulation results demonstrate that the proposed method can provide heterogeneous smart grid businesses with differentiated service, improve the utilization and economic benefits of the network and make the network more balanced.
Abstract: Network function virtualization (NFV), has been widely adopted in existing networks since it brings lower capital expenditure and operating expense, as well as flexible and elastic deployment. Based on NFV, dynamic Service function chain (SFC) provides more comprehensive and thorough traffic steering over a series of middleboxes, e.g., DPI, and IDS. Nevertheless, dynamic SFC involves internal state migration in middleboxes, making it hard to allocate SFC requests. We propose a reconfigurable SFC scheduling approach with consideration on resource constraints and race of the state migration to collaborate the execution of the SFC reconfiguration, in the reasonable fashion based on a heuristic algorithm. The evaluation is conducted through discrete event simulation to validate the efficiency of our reconfigurable SFC scheduling, and the results demonstrate that the proposed method outperforms First Come First Serve scheduling and random scheduling.
Abstract: In time-triggered ethernet (TTEthernet), designing and optimizing the static scheduling of TT messages to improve the real-time performance of the whole network control system is important. When there is a high load on TTEthernet network communication, the conventional static scheduling policy introduces certain problems, such as increased packet loss rate, load imbalance, and transmission delay. Meanwhile, basic artificial intelligence algorithms generally features convergence within only a few steps, leading to higher probability to fall into a local optimal solution. To realize these targets, Fuzzy particle swarm optimization (FPSO) based on population diversity is established. By combining Particle swarm optimization (PSO) with a fuzzy algorithm, setting the adjustment of the inertia weight and the regulation of the mutation factor as the controlled variables of the Fuzzy logic controller (FLC), and adjusting the FLC inference rules, a novel static schedule generation for TTEthernet is proposed. Simulation results prove that, compared to the conventional rate-monotonic scheduling algorithm and the PSO algorithm, FPSO has a strong global search capability when the communication of the TTEthernet system is in a high-load state. FPSO improves the load balancing of the network and reduces the transmission delay of TT message packets. FPSO shows excellent ability to optimize scheduling tables and guarantee the real-time performance of the TTEthernet system.
Abstract: A Reduced-dimensional spherical harmonics MUSIC (RD-SHMUSIC) is proposed to solve the problem of high computational complexity of Multiple signal classification (MUSIC) algorithm for Twodimensional (2-D) Direction of arrival (DOA) estimation. The proposed algorithm first expresses the spherical harmonic steering vector as a linear weight of a uniform phase vectors. Via the Lagrange multiplier method, we can get a new search function to estimate elevations. At second step, the algorithm expresses the steering vector in another form, a linear weight of a vector which is constructed by associated Legendre functions. A search function to estimate azimuths can also be obtained. The proposed method only needs One dimensional (1-D) angle search, which means it has a large reduction of computational complexity compared to the traditional Spherical harmonics MUSIC (SHMUSIC). The simulation results show that the accuracy of the proposed method is better than Two-stage decoupled approach (TSDA) algorithm, and has close performance to that of SHMUSIC.
Abstract: A coding metasurface is designed based on hybrid Array pattern synthesis (APS) and Particle swarm optimization (PSO) method for ultra-wideband low-detectable application. The metasurface is composed of Electromagnetic band-gap (EBG) structures of two Minkowski fractal elements with reflection phase difference of 180° (±37°) over a wide frequency range. Two different types of EBG unit cell printed on a thin grounded dielectric substrate produce reflection phase difference of about 180 degrees over a wide frequency range. Ultra-wideband Radar cross section (RCS) reduction results from the phase cancellation between two local waves produced by these two unit cells. The diffuse scattering of EM waves is caused by the optimization of phase distribution, leading to a low monostatic and bistatic RCS simultaneously. The proposed metasurface can achieve 10 dB monostatic and bistatic RCS reduction in a wide frequency band from 5.8 to 18.0 GHz with a ratio bandwidth (fH/fL) of 3.10:1 under normal incidence for both polarizations. The theoretical analysis, simulation and experiment results are in good agreement and validate the proposed metasurface can achieve ultra-wideband RCS reduction and diffuse scattering.
Abstract: In the problem of aircraft detection for High resolution (HR) Synthetic aperture radar (SAR) images, the background areas commonly contain multiple land cover types, such as runways and grassland. The conventional Constant false alarm rate (CFAR) detection in these non-homogeneous backgrounds with homogeneous assumption leads to unreliable detection results. This paper constructs a one-stage detection method based on the Generalized gamma mixture distribution (GGMD), which is regarded as a competitive and applicable model for combining the advantages of the Generalized gamma distribution (GGD) and the Finite mixture model (FMM). In order to evaluate the availability of the proposed algorithm, HR SAR images for aircraft detection from different product types and with various resolutions are examined. Compared with the CFAR algorithms based on the Gamma distribution, the GGD, and the gamma mixture distribution, the proposed algorithm demonstrates its availability and effectiveness for aircraft detection in HR SAR images in non-homogeneous background.