Two-stage Architecture Research Articles

Traffic Light Recognition for Real Scenes Based on Image Processing and Deep Learning

Traffic light recognition in urban environments is crucial for vehicle control. Many studies have been devoted to recognizing traffic lights. However, existing recognition methods still face many challenges in terms of accuracy, runtime and size. This paper presents a novel robust traffic light recognition approach that takes into account these three aspects based on image processing and deep learning. The proposed approach adopts a two-stage architecture, first performing detection and then classification. In the detection, the perspective relationship and the fractal dimension are both considered to dramatically reduce the number of invalid candidate boxes, i.e. region proposals. In the classification, the candidate boxes are classified by SqueezeNet. Finally, the recognized traffic light boxes are reshaped by postprocessing. Compared with several reference models, this approach is significantly competitive in terms of accuracy and runtime. We show that our approach is lightweight, easy to implement, and applicable to smart terminals, mobile devices or embedded devices in practice.

Ingredients to enhance the performance of two-stage TCAM-based packet classifiers in internet of things: greedy layering, bit auctioning and range encoding

Using packet classification algorithms in network equipment increases packet processing speed in Internet of Things (IoT). In the hardware implementation of these algorithms, ternary content-addressable memories (TCAMs) are often preferred to other implementations. As a common approach, TCAMs are used for the parallel search to match packet header information with the rules of the classifier. In two-stage architectures of hardware-based packet classifiers, first the decision tree is created, and then the rules are distributed among its leaves. In the second step, depending on the corresponding leaves, the second part of the rules, which includes the range of source and destination ports is stored in different blocks of TCAM. Due to inappropriate storage of port range fields, the existing architectures face the problem of wasting memory and growing power consumption. This paper proposes an efficient algorithm to encode the port range. This algorithm consists of three general steps including layering, bit allocation, and encoding. A greedy algorithm in the first step places the ranges with higher weights in higher layers. Next, an auction-based algorithm allocates several bits to each layer depending on the number of the ranges in that layer. Finally, in each layer, depending on the weight order of the ranges, the bits are given values for the intended range. The evaluation results show that unlike previous methods of storing range fields, the proposed method not only increases the speed of the classification but also uses the capacity of TCAM in the second stage more efficiently.

High-Frequency and High-Efficiency Isolated Two-Stage Bidirectional DC–DC Converter for Residential Energy Storage Systems

The main aim of this article is to develop an isolated bidirectional CLLC converter that achieves a wide output-voltage range for residential energy storage systems (ESS), while maintaining high efficiency and achieving high power density. The circuit is a two-stage architecture, where the first stage is an isolated CLLC converter with open-loop fixed-frequency control and the second stage is a buck converter used for output voltage regulation. The influence of the bus voltage on the efficiency of the system in both forward and reverse operation modes is analyzed. Then, an asymmetric resonant tank design is developed such that different voltage gains are achieved in each mode with high switching frequency operation at the resonance point to optimize efficiency. For high power density, an improved integrated transformer structure with controllable leakage is designed for the high step-down ratio specification of the ESS application. The core design is optimized based on core loss, copper loss, and core volume. A 1-kW experimental prototype using GaN switches was implemented with an input voltage of 380 V, output voltage of 35–50 V, highest efficiency of 94.55%, and power density of 5.56 W/cm3. The front-end stage CLLC converter operates at 700 kHz with a maximum efficiency of 97.02%.

Deep Neural Networks Improve Radiologists' Performance in Breast Cancer Screening.

We present a deep convolutional neural network for breast cancer screening exam classification, trained, and evaluated on over 200 000 exams (over 1 000 000 images). Our network achieves an AUC of 0.895 in predicting the presence of cancer in the breast, when tested on the screening population. We attribute the high accuracy to a few technical advances. 1) Our network’s novel two-stage architecture and training procedure, which allows us to use a high-capacity patch-level network to learn from pixel-level labels alongside a network learning from macroscopic breast-level labels. 2) A custom ResNet-based network used as a building block of our model, whose balance of depth and width is optimized for high-resolution medical images. 3) Pretraining the network on screening BI-RADS classification, a related task with more noisy labels. 4) Combining multiple input views in an optimal way among a number of possible choices. To validate our model, we conducted a reader study with 14 readers, each reading 720 screening mammogram exams, and show that our model is as accurate as experienced radiologists when presented with the same data. We also show that a hybrid model, averaging the probability of malignancy predicted by a radiologist with a prediction of our neural network, is more accurate than either of the two separately. To further understand our results, we conduct a thorough analysis of our network’s performance on different subpopulations of the screening population, the model’s design, training procedure, errors, and properties of its internal representations. Our best models are publicly available at https://github.com/nyukat/breast_cancer_classifier.

Real-Time Object Detection With Reduced Region Proposal Network via Multi-Feature Concatenation.

In recent years, object detection became more and more important following the successful results from studies in deep learning. Two types of neural network architectures are used for object detection: one-stage and two-stage. In this paper, we analyze a widely used two-stage architecture called Faster R-CNN to improve the inference time and achieve real-time object detection without compromising on accuracy. To increase the computation efficiency, pruning is first adopted to reduce the weights in convolutional and fully connected (FC) layers. However, this reduces the accuracy of detection. To address this loss in accuracy, we propose a reduced region proposal network (RRPN) with dilated convolution and concatenation of multi-scale features. In the assisted multi-feature concatenation, we propose the intra-layer concatenation and proposal refinement to efficiently integrate the feature maps from different convolutional layers; this is then provided as an input to the RRPN. Using the proposed method, the network can find object bounding boxes more accurately, thus compensating for the loss arising from compression. Finally, we test the proposed architecture using ZF-Net and VGG16 as a backbone network on the image sets in PASCAL VOC 2007 or VOC 2012. The results show that we can compress the parameters of the ZF-Net-based network by 81.2% and save 66% of computation. The parameters of VGG16-based network are compressed by 73% and save 77% of computation. Consequently, the inference speed is improved from 27 to 40 frames/s for ZF-Net and 9 to 27 frames/s for VGG16. Despite significant compression rates, the accuracy of ZF-Net is increased from 2.2% to 60.2% mean average precision (mAP) and that of VGG16 is increased from 2.6% to 69.1% mAP.

Deep Learning for Talker-dependent Reverberant Speaker Separation: An Empirical Study.

Speaker separation refers to the problem of separating speech signals from a mixture of simultaneous speakers. Previous studies are limited to addressing the speaker separation problem in anechoic conditions. This paper addresses the problem of talker-dependent speaker separation in reverberant conditions, which are characteristic of real-world environments. We employ recurrent neural networks with bidirectional long short-term memory (BLSTM) to separate and dereverberate the target speech signal. We propose two-stage networks to effectively deal with both speaker separation and speech dereverberation. In the two-stage model, the first stage separates and dereverberates two-talker mixtures and the second stage further enhances the separated target signal. We have extensively evaluated the two-stage architecture, and our empirical results demonstrate large improvements over unprocessed mixtures and clear performance gain over single-stage networks in a wide range of target-to-interferer ratios and reverberation times in simulated as well as recorded rooms. Moreover, we show that time-frequency masking yields better performance than spectral mapping for reverberant speaker separation.

Solving Current Limitations of Deep Learning Based Approaches for Plant Disease Detection

Plant diseases cause great damage in agriculture, resulting in significant yield losses. The recent expansion of deep learning methods has found its application in plant disease detection, offering a robust tool with highly accurate results. The current limitations and shortcomings of existing plant disease detection models are presented and discussed in this paper. Furthermore, a new dataset containing 79,265 images was introduced with the aim to become the largest dataset containing leaf images. Images were taken in various weather conditions, at different angles, and daylight hours with an inconsistent background mimicking practical situations. Two approaches were used to augment the number of images in the dataset: traditional augmentation methods and state-of-the-art style generative adversarial networks. Several experiments were conducted to test the impact of training in a controlled environment and usage in real-life situations to accurately identify plant diseases in a complex background and in various conditions including the detection of multiple diseases in a single leaf. Finally, a novel two-stage architecture of a neural network was proposed for plant disease classification focused on a real environment. The trained model achieved an accuracy of 93.67%.

Privacy-preserving deep learning algorithm for big personal data analysis

For privacy-preserving analysing of big data, a deep learning method is proposed. The method transforms the sensitive part of the personal information into non-sensitive data. To implement this process, two-stage architecture is proposed. The modified sparse denoising autoencoder and CNN models have been used in the architecture. Modified sparse denoising autoencoder performs transformation of data and CNN classifies the transformed data. In order to achieve low loss in data transformation, the sparsification parameter is added to the objective function of the autoencoder by the Kullback–Leibler divergence function. Here, the efficiency evaluation of the model is conducted by the MSE (mean squared error) loss function. In order to evaluate the accuracy of the transformation process, the features derived from the sparse denoising autoencoder algorithm fed to the input of the deep CNN algorithm and the classification of the reconstructed data is classified to the Black (0), White (1) and Gray (2) classes. Since here conducted the transformation of the Black class data to the Gray class data, in the classification stage, the CNN algorithm is classified the Black class data as the Gray class with 0.99 accuracy. The comparison of the proposed method with simple autoencoder is provided and experiments conducted on Cleveland medical dataset extracted from the Heart Disease dataset, Arrhythmia and Skoda datasets showed that the proposed method outperforms other conventional methods.

Programmable Analog Calculation Unit with Two-Stage Architecture: A Solution of Efficient Vector-Computation

Programmable Analog Calculation Unit with Two-Stage Architecture: A Solution of Efficient Vector-Computation

On the adaptability of ensemble methods for distributed classification systems: A comparative analysis

In this work, a two-stage architecture is used to analyze the information collected from several sensors. The first stage makes classifications from partial information of the entire target (i.e. from different points of view or from different kind of measures) using a simple artificial neural network as a classifier. In addition, the second stage aggregates all the estimations given by the ensemble in order to obtain the final classification. Four different ensembles methods are compared in the second stage: artificial neural network, plurality majority, basic weighted majority, and stochastic weighted majority. However, not only reliability is an important factor but also adaptation is critical when the ensemble is working in changing environments. Therefore, the artificial neural network and the plurality majority algorithm are compared against our two proposed adaptive algorithms. Unlike artificial neural network, majority methods do not require previous training. The effects of improving the first stage and how the system behaves when different perturbations are presented have been measured. Results have been obtained from two applications: a realistic one and another simpler one, with more training examples for a more accurate comparison. These results show that artificial neural network is the most accurate proposal, whereas the most innovative proposed stochastic weighted voting is the most adaptive one.

Proximal Support Vector Machine-Based Hybrid Approach for Edge Detection in Noisy Images

Abstract We propose a novel edge detector in the presence of Gaussian noise with the use of proximal support vector machine (PSVM). The edges of a noisy image are detected using a two-stage architecture: smoothing of image is first performed using regularized anisotropic diffusion, followed by the classification using PSVM, termed as regularized anisotropic diffusion-based PSVM (RAD-PSVM) method. In this process, a feature vector is formed for a pixel using the denoised coefficient’s class and the local orientations to detect edges in all possible directions in images. From the experiments, conducted on both synthetic and benchmark images, it is observed that our RAD-PSVM approach outperforms the other state-of-the-art edge detection approaches, both qualitatively and quantitatively.

An intelligent intrusion detection system

With the introduction of emerging technologies cybersecurity has become an inherited and amplified problem. New technologies bring significant developments but also come with new challenges in the cybersecurity area. The fight against malicious attacks is an everyday battle for every company. Challenges brought by security breaches can be devastating for a company and sometimes bring un-survivable circumstances. In this paper, we propose a novel two-stage intelligent intrusion detection system (IDS) to detect and protect from such malicious attacks. Intrusion Detection Systems are feasible solutions for cybersecurity problems, but they come with implementation challenges. Anomaly based IDS usually have a high rate of false positives (FP) and they require considerable computational requirements. The approach proposed in this paper consists of a two-stage architecture based on machine learning algorithms. In the first stage, the IDS uses K-Means to detect attacks and the second stage uses supervised learning to classify such attacks and eliminate the number of false positives. The implementation of this approach results in a computationally efficient IDS able to detect and classify attacks at a 99.97% accuracy while lowering the number of false positives to 0. The paper also evaluates the performance results and compares them with other relevant research papers. The performance of this proposed IDS is superior to the current state of the art.

A Two-Stage Switched-Capacitor Integrator for High Gain Inverter-Like Architectures

A discrete-time, switched capacitor integrator is presented. The integrator is based on a two-stage architecture where the first stage converts the input voltage into a charge that is accumulated into the second stage. The main strength of the proposed circuit is a higher dc gain with respect to previous solutions, making it optimal for low-voltage inverter-like integrators. A further advantage is the fact that, in contrast with existing solutions, the output voltage is valid across the whole clock cycle. Theoretical analysis of the circuit is performed to calculate the dependence of the integrator dc gain and input-referred offset voltage on the corresponding parameters of the constituting amplifiers. Discrete-time simulations are performed to estimate the gain and phase error with respect to an ideal integrator. The results of electrical simulations performed on an inverter-like prototype, designed with the UMC 0.18- ${\mu }\text{m}$ CMOS process, are presented to show the impact of non-idealities from the amplifiers and switches.

A Two-Stage Kalman Filter-Based Carrier Tracking Loop for Weak GNSS Signals.

For global navigation satellite system receivers, Kalman filter (KF)-based tracking loops show remarkable advantages in terms of tracking sensitivity and robustness compared with conventional tracking loops. However, to improve the tracking sensitivity further, increasing the coherent integration time is necessary, but it is typically limited by the navigation data bit sign transition. Moreover, for standard KF-based tracking receivers, the KF parameters are initialized by the acquired results. However, especially under weak signal conditions, the acquired results have frequency errors that are too large for KF-based tracking to converge rapidly to a steady state. To solve these problems, a two-stage KF-based tracking architecture is proposed to track weaker signals and achieve faster convergence. In the first stage, coarse tracking refines the acquired results and achieves bit synchronization. Then, in the second stage, fine tracking initializes the KF-based tracking by using the coarse tracking results and extends the coherent integration time without the bit sign transition limitation. This architecture not only utilizes the self-tuning technique of the KF to improve the tracking sensitivity, but also adopts the two-stage to reduce the convergence time of the KF-based tracking. Simulation results demonstrate that the proposed method outperforms conventional tracking techniques in terms of tracking sensitivity. Furthermore, the proposed method is compared with the standard KF-based tracking approach, proving that the proposed method converges more rapidly.

Mixed-technology quasi-reflectionless planar filters: bandpass, bandstop, and multi-band designs

AbstractThe design of mixed-technology quasi-reflectionless planar bandpass filters (BPFs), bandstop filters (BSFs), and multi-band filters is reported. The proposed quasi-reflectionless filter architectures comprise a main filtering section that determines the power transmission response (bandpass, bandstop, or multi-band type) of the overall circuit network and auxiliary sections that absorb the reflected radio-frequency (RF) signal energy. By loading the input and output ports of the main filtering section with auxiliary filtering sections that exhibit a complementary transfer function with regard to the main one, a symmetric quasi-reflectionless behavior can be obtained at both accesses of the overall filter. The operating principles of the proposed filter concept are shown through synthesized first-order BPF and BSF designs. Selectivity-increase techniques are also described. They are based on: (i) cascading in-series multiple first-order stages and (ii) increasing the order of the filtering sections. Moreover, the RF design of quasi-reflectionless multi-band BPFs and BSFs is discussed. A hybrid integration scheme in which microstrip-type and lumped-elements are effectively combined within the filter volume is investigated for size miniaturization purposes. For experimental validation purposes, two quasi-reflectionless BPF prototypes (one- and two-stage architectures) centered at 2 GHz and a second-order BSF prototype centered at 1 GHz were designed, manufactured, and measured.

Postnatal gestational age estimation of newborns using Small Sample Deep Learning

A baby's gestational age determines whether or not they are premature, which helps clinicians decide on suitable post-natal treatment. The most accurate dating methods use Ultrasound Scan (USS) machines, but these are expensive, require trained personnel and cannot always be deployed to remote areas. In the absence of USS, the Ballard Score, a postnatal clinical examination, can be used. However, this method is highly subjective and results vary widely depending on the experience of the examiner. Our main contribution is a novel system for automatic postnatal gestational age estimation using small sets of images of a newborn's face, foot and ear. Our two-stage architecture makes the most out of Convolutional Neural Networks trained on small sets of images to predict broad classes of gestational age, and then fuses the outputs of these discrete classes with a baby's weight to make fine-grained predictions of gestational age using Support Vector Regression. On a purpose-collected dataset of 130 babies, experiments show that our approach surpasses current automatic state-of-the-art postnatal methods and attains an expected error of 6 days. It is three times more accurate than the Ballard method. Making use of images improves predictions by 33% compared to using weight only. This indicates that even with a very small set of data, our method is a viable candidate for postnatal gestational age estimation in areas were USS is not available.

Two-Stage Convolutional Neural Network for Ship and Spill Detection Using SLAR Images

This paper presents a system for the detection of ships and oil spills using side-looking airborne radar (SLAR) images. The proposed method employs a two-stage architecture composed of three pairs of convolutional neural networks (CNNs). Each pair of networks is trained to recognize a single class (ship, oil spill, and coast) by following two steps: a first network performs a coarse detection, and then, a second specialized CNN obtains the precise localization of the pixels belonging to each class. After classification, a postprocessing stage is performed by applying a morphological opening filter in order to eliminate small look-alikes, and removing those oil spills and ships that are surrounded by a minimum amount of coast. Data augmentation is performed to increase the number of samples, owing to the difficulty involved in obtaining a sufficient number of correctly labeled SLAR images. The proposed method is evaluated and compared to a single multiclass CNN architecture and to previous state-of-the-art methods using accuracy, precision, recall, F-measure, and intersection over union. The results show that the proposed method is efficient and competitive, and outperforms the approaches previously used for this task.

Optimal Scheduling of a Microgrid Including Pump Scheduling and Network Constraints

This paper proposes an efficient energy management system (EMS) for industrial microgrids (MGs). Many industries deploy large pumps for their processes. Oftentimes, such pumps are operated during hours of peak electricity prices. A lot of industries use a mix of captive generation and imported utility electricity to meet their energy requirements. The MG considered in this paper includes diesel generators, battery energy storage systems, renewable energy sources, flexible loads, and interruptible loads. Pump loads found in shipyard dry docks are modelled as exemplar flexible industrial loads. The proposed EMS has a two-stage architecture. An optimal MG scheduling problem including pump scheduling and curtailment of interruptible loads (ILs) is formulated and solved in the first stage. An optimal power flow problem is solved in the second stage to verify the feasibility of the MG schedule with the network constraints. An iterative procedure is used to coordinate the two EMS stages. Multiple case studies are used to demonstrate the utility of the proposed EMS. The case studies highlight the efficacy of load management strategies such as pump scheduling and curtailment of ILs in reducing the total electricity cost of the MG.

A compact current-mode instrumentation amplifier for general-purpose sensor interfaces

The proposed amplifier architecture follows a consolidated topology based on second-generation current conveyors (CCIIs), optimized for fully-differential operation. The architecture uses gain-boosting to improve the offset and noise characteristics of a recently proposed design. Wide input and output ranges and high accuracy are obtained by designing the CCIIs according to an original two-stage architecture with local voltage feedback. Embedding of chopper switch matrices into the amplifier enables vector analysis of the input signal, expanding the application field. The main strengths of the proposed amplifier are compactness and versatility. Measurements performed on a prototype designed with a 0.18 μm CMOS process are described.

Building a model-based personalised recommendation approach for tourist attractions from geotagged social media data

ABSTRACTWhen travelling, people are accustomed to taking and uploading photos on social media websites, which has led to the accumulation of huge numbers of geotagged photos. Combined with multisource information (e.g. weather, transportation, or textual information), these geotagged photos could help us in constructing user preference profiles at a high level of detail. Therefore, using these geotagged photos, we built a personalised recommendation system to provide attraction recommendations that match a user's preferences. Specifically, we retrieved a geotagged photo collection from the public API for Flickr (Flickr.com) and fetched a large amount of other contextual information to rebuild a user's travel history. We then created a model-based recommendation method with a two-stage architecture that consists of candidate generation (the matching process) and candidate ranking. In the matching process, we used a support vector machine model that was modified for multiclass classification to generate the candidate list. In addition, we used a gradient boosting regression tree to score each candidate and rerank the list. Finally, we evaluated our recommendation results with respect to accuracy and ranking ability. Compared with widely used memory-based methods, our proposed method performs significantly better in the cold-start situation and when mining ‘long-tail’ data.