Intensive Computational Complexity Research Articles

On the search for efficient face recognition algorithm subject to multiple environmental constraints

From literature, majority of face recognition modules suffer performance challenges when presented with test images acquired under multiple constrained environments (occlusion and varying expressions). The performance of these models further deteriorates as the degree of degradation of the test images increases (relatively higher occlusion level). Deep learning-based face recognition models have attracted much attention in the research community as they are purported to outperform the classical PCA-based methods. Unfortunately their application to real-life problems is limited because of their intensive computational complexity and relatively longer run-times. This study proposes an enhancement of some PCA-based methods (with relatively lower computational complexity and run-time) to overcome the challenges posed to the recognition module in the presence of multiple constraints. The study compared the performance of enhanced classical PCA-based method (HE-GC-DWT-PCA/SVD) to FaceNet algorithm (deep learning method) using expression variant face images artificially occluded at 30% and 40%. The study leveraged on two statistical imputation methods of MissForest and Multiple Imputation by Chained Equations (MICE) for occlusion recovery. From the numerical evaluation results, although the two models achieved the same recognition rate (85.19%) at 30% level of occlusion, the enhanced PCA-based algorithm (HE-GC-DWT-PCA/SVD) outperformed the FaceNet model at 40% occlusion rate, with a recognition rate of 83.33%. Although both Missforest and MICE performed creditably well as de-occlusion mechanisms at higher levels of occlusion, MissForest outperforms the MICE imputation mechanism. MissForest imputation mechanism and the proposed HE-GC-DWT-PCA/SVD algorithm are recommended for occlusion recovery and recognition of multiple constrained test images respectively.

Polynomial chaos Kriging-based structural reliability analysis via the expected margin volume reduction

Integral learning functions are theoretically sound but suffer from intensive computational complexity induced by the associated double integral. Based on a new notation of limit-state margin volume, a computationally-cheap integral learning function called Expected margin volume reduction (EMVR) is proposed for structural reliability analysis. EMVR has two key contributions. First, closed-form expression for the inner integral is well derived based on Kriging update formulas, due to tractable definition of the margin volume. This gets rid of cumbersome numerical quadrature or drawing massive realizations of Gaussian process. Second, a confined integral domain is rationally defined for the outer integral, by virtue of exploring the locality of its integrand. This bypasses annoying computer memory issue. Moreover, a hybrid stopping condition coupled with two different settings of the associated parameters is deployed, accommodating to reliability problems with different features. Then, the performance of EMVR-based reliability algorithm is illustrated on four numerical examples. The results show that the evaluation time of EMVR is reduced to a level comparable to pointwise learning functions. Moreover, it outperforms those existing learning functions in terms of both computational accuracy and efficiency.

A 250-mW 5.4G-Rendered-Pixel/s Realistic Refocusing Processor for High-Performance Five-Camera Mobile Devices

Digital refocusing in multi-camera mobile devices is becoming crucial. Realistic refocusing, which is a subset of digital refocusing, provides physically-correct quality; however, its intense computational complexity results in low processing speed and restricts its applicability. Moreover, its complex computation flow requires substantial DRAM bandwidth and a large SRAM area, making it more challenging to implement in hardware. In this paper, we present a high-performance refocusing processor based on a hardware-oriented realistic refocusing algorithm. The proposed compact computation flow saves 92% of the DRAM bandwidth and 32% of the SRAM area without noticeable quality degradation. To support high-performance refocusing, we develop highly paralleled engines for view rendering. They deliver 5.4G rendered-pixel/sec throughput. The hardware accelerator improves the processing speed by 100× to 350× that of the original refocusing algorithm running on a general purpose processor. The chip is fabricated with 40nm CMOS technology and comprises 271KB of SRAM and 2.3M logic gates. The chip processes Full-HD light fields up to 40 frames per second under 250mW power consumption.

A power-aware ECG processing node for real-time feature extraction in WBAN

In modern WBAN (Wireless Body Area Network) applications, wearable sensor nodes are advancing towards in-node detection of cardiovascular diseases with continuous observation of ECG signal. However, a robust power-aware architecture with in-node processing for extraction of cardiac events triggers challenges due to constraints in hardware resources in addition to intensive computational complexity. In this work, a dual mode based transmission system depending on battery-level is discussed that reduces bitstream load of raw ECG signal for saving transmission power. For encoding the bitstream, a power-aware ECG processing architecture is proposed that performs a real-time feature extraction for continuous health monitoring. The proposed architecture explores the advantage of low computationally complex operations focusing on a resource constraint sensor node. The architecture deals with two types of cardiac anomalies i.e., irregular heart rate variation (HRV) and 1st degree AV block; detecting R-peak and P-peak through adaptive thresholding technique. The overall sensitivity and positive predictivity for the proposed algorithm reach upto 99% while validated with the MIT-BIH arrhythmia database. The proposed architecture is synthesized in TSMC 90 nm technology node with an estimated power consumption of 2.11 µW, consuming an area overhead of 0.087 mm2. The architecture operates at 1.2 V with an operating frequency of 10 KHz.

Machine Learning for Radio Resource Management in Multibeam GEO Satellite Systems

Satellite communications (SatComs) systems are facing a massive increase in traffic demand. However, this increase is not uniform across the service area due to the uneven distribution of users and changes in traffic demand diurnal. This problem is addressed by using flexible payload architectures, which allow payload resources to be flexibly allocated to meet the traffic demand of each beam. While optimization-based radio resource management (RRM) has shown significant performance gains, its intense computational complexity limits its practical implementation in real systems. In this paper, we discuss the architecture, implementation and applications of Machine Learning (ML) for resource management in multibeam GEO satellite systems. We mainly focus on two systems, one with power, bandwidth, and/or beamwidth flexibility, and the second with time flexibility, i.e., beam hopping. We analyze and compare different ML techniques that have been proposed for these architectures, emphasizing the use of Supervised Learning (SL) and Reinforcement Learning (RL). To this end, we define whether training should be conducted online or offline based on the characteristics and requirements of each proposed ML technique and discuss the most appropriate system architecture and the advantages and disadvantages of each approach.

Easy Domain Adaptation for cross-subject multi-view emotion recognition

Existing domain adaptation methods for cross-subject emotion recognition are primarily focused on accuracy and suffer from the issues of intensive hyperparameter tunings and high computational complexity. In this paper, we make the first attempt to address these issues by developing a domain-invariant classifier called Easy Domain Adaptation (EasyDA) based on multi-view emotion inputs (multiple modalities or multiple types of features). Firstly, EasyDA uses both the source domain (training subjects) and the target domain (test subject) to generate domain generalization features for each view by leveraging a fast, accurate, and low-memory approximate empirical kernel map (AEKM), followed by a parameterless weighted combination for multi-views. Secondly, EasyDA simultaneously learns an optimal separating hyperplane and the pseudo labels for the target domain such that (a) high classification accuracies are obtained on both labeled source domain data and the pseudo-labeled target domain data; (b) the distribution distance between source and target domain is reduced; (c) the predicted output vector in the target domain changes little over time in short time intervals, based on the biological evidence that emotion varies fluently and smoothly. Eventually, by summarizing these two steps with the ridge regression theory and alternating optimization, EasyDA can transfer knowledge across domains accurately, efficiently, and easily in a unified framework. Experimental results on the SEED and SEED-IV datasets demonstrate that EasyDA significantly outperforms multiple representative domain adaptation methods in terms of accuracy, computational time, and memory consumption. It is noteworthy that EasyDA achieves satisfactory performance under a wide range of parameter settings.

A Low-Complexity Hyperspectral Anomaly Detection Algorithm and Its FPGA Implementation

On-board real-time anomaly detection has always been a challenging task in hyperspectral imaging analysis as it requires low computational complexity. Most of the existing anomaly detection algorithms inevitably trade off intensive computational complexity for high detection accuracy. This article presents a fast spectral-spatial anomaly detection algorithm with low complexity in hyperspectral images (HSIs) using morphological reconstruction and a simplified guided filter (Fast-MGD). Since the simple filtering techniques are applied, it is therefore feasible to achieve a field programmable gate array (FPGA)-based hardware implementation. More precisely, an effective deeply pipelined acceleration scheme is developed adopting high-level synthesis to support HSIs that are acquired over different scenes with different sizes and spectral bands. Experimental results show strong advantages of the proposed FPGA-based Fast-MGD in processing speed and resource consumption, while a high detection accuracy is remained. Its applicability in on-board real-time processing is demonstrated and verifie.

Saliency Detection via the Improved Hierarchical Principal Component Analysis Method

Aiming at the problems of intensive background noise, low accuracy, and high computational complexity of the current significant object detection methods, the visual saliency detection algorithm based on Hierarchical Principal Component Analysis (HPCA) has been proposed in the paper. Firstly, the original RGB image has been converted to a grayscale image, and the original grayscale image has been divided into eight layers by the bit surface stratification technique. Each image layer contains significant object information matching the layer image features. Secondly, taking the color structure of the original image as the reference image, the grayscale image is reassigned by the grayscale color conversion method, so that the layered image not only reflects the original structural features but also effectively preserves the color feature of the original image. Thirdly, the Principal Component Analysis (PCA) has been performed on the layered image to obtain the structural difference characteristics and color difference characteristics of each layer of the image in the principal component direction. Fourthly, two features are integrated to get the saliency map with high robustness and to further refine our results; the known priors have been incorporated on image organization, which can place the subject of the photograph near the center of the image. Finally, the entropy calculation has been used to determine the optimal image from the layered saliency map; the optimal map has the least background information and most prominently saliency objects than others. The object detection results of the proposed model are closer to the ground truth and take advantages of performance parameters including precision rate (PRE), recall rate (REC), and F-measure (FME). The HPCA model’s conclusion can obviously reduce the interference of redundant information and effectively separate the saliency object from the background. At the same time, it had more improved detection accuracy than others.

A Spatiotemporal Content-Based CU Size Decision Algorithm for HEVC

The high efficiency video coding (HEVC) standard provides superior efficiency for encoding and improves the compression ratio by almost 50% compared with previous video coding standards, such as advanced video coding (AVC). However, more intensive computation complexity is introduced by implementing the flexible quad tree-structured coding model. In a typical HEVC encoder, coding units (CUs) in a coding tree unit (CTU) that is built as a quad-tree structure are recursively traversed each depth level (CU size) to select the optimal coding configuration. Therefore, most of the encoding time is spent searching for the optimal coding configuration. In this paper, an efficient and fast CU size decision algorithm is proposed to reduce HEVC encoder complexity by the spatiotemporal features. First, an adaptive depth-range prediction method minimizes the possible range depth level by observing previous frames and proximal CTUs. Second, an early termination method based on the boundary examination from the de-blocking filter (DBF) prevents unnecessary calculation on small CU sizes. Furthermore, according to the sum of absolution difference (SAD), a smooth area detection mechanism is triggered when the predictive depth range excludes the largest CU size. This mechanism increase the bitrate of the CU, which contains static objects with complex textures. Compared with the HM 16, the experimental results revealed that the proposed algorithm can achieve an average 59.73% and 64.98% reduction in encoding time along with a 0.68% and 1.27% Bjontegaard Delta bitrate (BDBR) penalty for various test videos under low-delay P and random-access conditions, respectively.

Multi-View Spectral Clustering with High-Order Optimal Neighborhood Laplacian Matrix

Multi-view spectral clustering can effectively reveal the intrinsic cluster structure among data by performing clustering on the learned optimal embedding across views. Though demonstrating promising performance in various applications, most of existing methods usually linearly combine a group of pre-specified first-order Laplacian matrices to construct the optimal Laplacian matrix, which may result in limited representation capability and insufficient information exploitation. Also, storing and implementing complex operations on the {$n\times n}$ Laplacian matrices incurs intensive storage and computation complexity. To address these issues, this paper first proposes a multi-view spectral clustering algorithm that learns a high-order optimal neighborhood Laplacian matrix, and then extends it to the late fusion version for accurate and efficient multi-view clustering. Specifically, our proposed algorithm generates the optimal Laplacian matrix by searching the neighborhood of the linear combination of both the first-order and high-order base Laplacian matrices simultaneously. By this way, the representative capacity of the learned optimal Laplacian matrix is enhanced, which is helpful to better utilize the hidden high-order connection information among data, leading to improved clustering performance. We design an efficient algorithm with proved convergence to solve the resultant optimization problem. Extensive experimental results on nine datasets demonstrate the superiority of the proposed algorithm

Deep Learning Approach Based on Tensor-Train for Sparse Signal Recovery

Compressive sensing is a desirable technique to acquire and reconstruct signals at sub-Nyquist rates. Recently, several deep learning-based studies on solving the compressive sensing problem have been carried out, which dramatically reduce the intensive computational complexity of the traditional greedy or convex recovery algorithms and even improve the signal recovery performance. However, as the signal size increases, most of these methods recover signals block by block due to the large computational complexity and memory consumption, which usually imposes block effect on the recovered signals. To deal with this issue, in this paper, we apply a tensor decomposition method named Tensor-Train (TT) on the neural network, by which the number of parameters is reduced by a tremendous factor and the computational complexity is further decreased so that the large signals can be recovered as a whole. In particular, the TT-decomposition is jointly applied on a stacked denoising autoencoder (SDA) network called TT-SDA in this paper. The experiments indicate that the proposed TT-SDA network can preserve the reconstruction performance of the conventional SDA network and outperform the traditional methods, especially with low measurement rates. Meanwhile, it can also significantly reduce the computational complexity and occupied memory space, which becomes a time and memory efficient method in compressive sensing problem.

Blind CFO Estimations Based on Subspace-Based Approaches Under Single Data Block for Interleaved OFDMA Uplink

This letter presents computationally efficient subspace-based estimators to estimate carrier frequency offset (CFO) for interleaved orthogonal frequency division multiple access (OFDMA) uplink systems. Conventional subspace-based estimators rely on eigenvalue decomposition (EVD) of sample correlation matrix to compute signal subspace or noise subspace. For interleaved OFDMA with large subchannel, conventional subspace-based CFO estimators inevitably lead to intensive computational complexity due to calculating sample correlation matrix and its EVD. To circumvent this problem, in conjugation with center-symmetric trimmed correlation matrix and Nystrom method, the required signal subspace or orthogonal projection matrix only needs to calculate two submatrices of center-symmetric trimmed correlation matrix. Therefore, the proposed signal subspace-based and noise subspace-based estimators not only can estimate CFO, but also save computational cost, especially in high SNR and large subchannel scenarios. Computer simulation results are also provided for illustrating the effectiveness of the presented blind CFO estimators under single OFDMA data block.

A Cancelable Fuzzy Vault Algorithm Based on Transformed Fingerprint Features

Juels and Sudan proposed in 2002 an algorithm for computing a fuzzy vault that binds a user's biometric template with his secret. It was suggested that this vault could be used to securely store one's secret or a cryptographic key without losing his biometric information. However, in this classical fuzzy vault, if an attacker captures multiple vaults generated from one biometric template, he is able to obtain some biometric template information by cross-matching, and then he can use it to illegally recover the secret message. To overcome this disadvantage, in this paper, a cancelable fuzzy vault algorithm is proposed based on the user's transformed fingerprint features which are used to generate a fuzzy vault. Our novel fuzzy vault is secure and can overcome the cross-matching attack without intensive computational complexity. Also, the use of three check values makes our vault have a much higher probability to detect a false query fingerprint template than some other vault versions, and it will highly improve the probability whether the reconstructed polynomial is correct or not.

ESPACE: Accelerating Convolutional Neural Networks via Eliminating Spatial and Channel Redundancy

Recent years have witnessed an extensive popularity of convolutional neural networks (CNNs) in various computer vision and artificial intelligence applications. However, the performance gains have come at a cost of substantially intensive computation complexity, which prohibits its usage inresource-limited applications like mobile or embedded devices. While increasing attention has been paid to the acceleration of internal network structure, the redundancy of visual input is rarely considered. In this paper, we make the first attempt of reducing spatial and channel redundancy directly from the visual input for CNNs acceleration. The proposed method, termed ESPACE (Elimination of SPAtial and Channel rEdundancy), works by the following three steps: First, the 3D channel redundancy of convolutional layers is reduced by a set of low-rank approximation of convolutional filters. Second, a novel mask based selective processing scheme is proposed, which further speedups the convolution operations via skipping unsalient spatial locations of the visual input. Third, the accelerated network is fine-tuned using the training data via back-propagation. The proposed method is evaluated on ImageNet 2012 with implementations on two widely adopted CNNs, i.e. AlexNet and GoogLeNet. In comparison to several recent methods of CNN acceleration, the proposed scheme has demonstrated new state-of-the-art acceleration performance by a factor of 5.48* and 4.12* speedup on AlexNet and GoogLeNet, respectively, with a minimal decrease in classification accuracy.

Image super-resolution using non-local Gaussian process regression

Gaussian process regression (GPR) is an elegant statistical learning method for nonlinear mappings. Despite its effectiveness, GPR suffers from intensive computational complexity. To improve the efficiency, most existing GPR-based super-resolution (SR) methods are local patch-based, which cannot make full use of self-similarities existing in natural images. To address this problem, we propose a non-local self-learning SR framework based on GPR, which learns only one non-local GPR model instead of multiple local patch-based models for SR reconstruction. In the proposed framework, an anisotropic linear kernel is introduced to construct a new kernel function for capturing more structure similarity. Furthermore, we disclose that a simple grid patch sampling with moderate sampling interval can be used to speed up the SR processing significantly without compromise of reconstruction quality. In addition, we make a theoretical discussion on two important factors of the predictive variance and the condition number of the kernel matrix that relate to the performance of the GPR. Experimental results on the benchmark test images show that our proposed method is superior to other state-of-art competitors in terms of both quantitative and qualitative measurements.

Low complexity encoder optimization for HEVC

High Efficiency Video Coding (HEVC) improved the coding efficiency significantly. Compared to its predecessor H.264/AVC, it can provide equivalent subjective quality with more than 57% bit rate reduction. However, the improvement on coding efficiency is obtained at the expense of much more intensive computation complexity. In this paper, based on an overall analysis of computation complexity at the HEVC encoder, a low complexity encoder optimization scheme is proposed by reducing the number of available candidates for evaluation in terms of the intra prediction mode decision, early termination of coding unit (CU) splitting and adaptive reference frame selection. With the proposed scheme, the rate distortion optimization (RDO) technique of HEVC can be implemented in a low-complexity way for complexity-constrained encoders. Experimental results demonstrate that, compared with the original HEVC reference encoder implementation, the proposed optimization scheme can achieve more than 40% complexity reduction on average with coding performance degradation as only 0.43% which can be ignorable.

Fast HEVC Inter CU Decision Based on Latent SAD Estimation

The emerging high efficiency video coding (HEVC) standard has improved compression performance significantly in comparison with H.264/AVC. However, more intensive computational complexity has been introduced by adopting a number of new coding tools. In this paper, a fast inter CU decision is proposed based on the latent sum of absolute differences (SAD) estimation. Firstly, a two-layer motion estimation (ME) method is designed to take advantage of the latent SAD cost. The new ME method can obtain the SAD costs for both the upper CU and its sub-CUs. Secondly, a concept of motion compensation rate- distortion (R-D) cost is defined, and an exponential model is proposed to express the relationship between the motion compensation R-D cost and the SAD cost. Then, a fast CU decision approach is designed based on the exponential model. The fast CU decision is implemented by comparing a derived threshold with the SAD cost difference between the upper and sub SAD costs. Experimental results show that the proposed algorithm achieves an average of 52% and 58.4% reductions of the coding time at the cost of 1.61% and 2% bit-rate increases under the low delay and random access conditions, respectively.

An Efficient VLSI Architecture of the Enhanced Three Step Search Algorithm

The intense computational complexity of any video codec is largely due to the motion estimation unit. The Enhanced Three Step Search is a popular technique that can be adopted for fast motion estimation. This paper proposes a novel VLSI architecture for the implementation of the Enhanced Three Step Search Technique. A new addressing mechanism has been introduced which enhances the speed of operation and reduces the area requirements. The proposed architecture when implemented in Verilog HDL on Virtex-5 Technology and synthesized using Xilinx ISE Design Suite 14.1 achieves a critical path delay of 4.8 ns while the area comes out to be 2.9K gate equivalent. It can be incorporated in commercial devices like smart-phones, camcorders, video conferencing systems etc.

MRF-Based Fast HEVC Inter CU Decision With the Variance of Absolute Differences

The newly developed High Efficiency Video Coding (HEVC) Standard has improved video coding performance significantly in comparison to its predecessors. However, more intensive computation complexity is introduced by implementing a number of new coding tools. In this paper, a fast coding unit (CU) decision based on Markov random field (MRF) is proposed for HEVC inter frames. First, it is observed that the variance of the absolute difference (VAD) is proportional with the rate-distortion (R-D) cost. The VAD based feature is designed for the CU selection. Second, the decision of CU splittings is modeled as an MRF inference problem, which can be optimized by the Graphcut algorithm. Third, a maximum a posteriori (MAP) approach based on the R-D cost is conducted to evaluate whether the unsplit CUs should be further split or not. Experimental results show that the proposed algorithm can achieve about 53% reduction of the coding time with negligible coding performance degradation, which outperforms the state-of-the-art algorithms significantly.

Fast reduced-rank STAP algorithm based on Gram–Schmidt orthogonalisation for airborne radar

In the reduced-rank space-time adaptive processing (STAP) methods, especially the principal component (PC) analysis STAP method, a set of dominant eigenvectors must be obtained by singular value decomposition of the space-time covariance matrix. Therefore, it is very difficult to be applied in practical system due to the intense computational complexity. In order to reduce the computational burden, a fast reduced-rank STAP algorithm based on Gram–Schmidt (GS) orthogonalisation is proposed in this article. In the proposed GS-PC STAP method, the clutter subspace is reconstructed by the GS orthogonalisation of training samples. Then, the STAP adaptive weight vector is calculated by orthogonally projecting the quiescent weight vector into clutter subspace, which can hold fast convergence measure of effectiveness (MOE) and require less computational complexity by compared with the conventional PC method. Based on the simulated data and multichannel airborne radar measurements data, the corresponding convergence MOE and the clutter suppression performances are verified in the article.