Search

With the evolution of electronic devices, such as 3D cameras, addressing the challenges of text localization in 3D video (e.g., for indexing) is increasingly drawing the attention of the multimedia and video processing community. Existing methods focus on 2D video and their performance in the presence of the challenges in 3D video, such as shadow areas associated with text and irregularly sized and shaped text, degrades. This paper proposes the first approach that successfully addresses the challenges of 3D video in addition to those of 2D. It employs a number of innovations, among which, the first is the Generalized Gradient Vector Flow (GGVF) for dominant points detection. The second is the Wavefront concept for text candidate point detection from those dominant points. In addition, an Adaptive B-Spline Polygon Curve Network (ABS-Net) is proposed for accurate text localization in 3D videos by constructing tight fitting bounding polygons using text candidate points. Extensive experiments on custom (3D video) and standard datasets (2D video and scene text) show that the proposed method is practical and useful, and overall outperforms existing state-of-the-art methods.

This paper focuses on automatic Cholangiocarcinoma (CC) diagnosis from microscopic hyperspectral (HSI) pathological dataset with deep learning method. The first benchmark based on the microscopic hyperspectral pathological images is set up. Particularly, 880 scenes of multidimensional hyperspectral Cholangiocarcinoma images are collected and manually labeled each pixel as either tumor or non-tumor for supervised learning. Moreover, each scene from the slide is given a binary label indicating whether it is from a patient or a normal person. Different from traditional RGB images, the HSI acquires pixels in multiple spectral intervals, which is added as an extension on the channel dimension of 3-channel RGB image. This work aims at fully exploiting the spatial-spectral HSI data through a deep Convolution Neural Network (CNN). The whole scene is first divided into several patches. Then they are fed into CNN for the tumor/non-tumor binary prediction and the tumor area regression. The further diagnosis on the scene is made by random forest based on the features from patch prediction. Experiments show that HSI provides a more accurate result than RGB image. Moreover, a spectral interval convolution and normalization scheme are proposed for further mining the spectral information in HSI, which demonstrates the effectiveness of the spatial-spectral data for CC diagnosis.

The 0‐order natural frequency resonance is the main cause of electromagnetic vibration noise in integer slot multi‐pole pair permanent magnet synchronous motors (PMSMs) for electric vehicles. To study this phenomenon, this study proposes a finite element (FE) model of the stator system based on a 48‐slot 8‐pole permanent magnet synchronous motor for electric vehicles. Based on this model, the effects of winding, frame end cover and temperature on the natural frequencies of a stator system are analysed by using the equivalent stiffness method (ESM). The simulation results show that the 0‐order natural frequency of the stator core system decreases after considering the winding varnished, and the 0‐order natural frequency is almost unaffected after considering the frame and end cover. However, the other natural frequencies increase in different degrees. In addition, the natural frequencies of the stator core decrease with the increase of temperature. The accuracy of the finite element model is verified by the modal experiment of the hammer method (HM).

With the growing needs of practical applications such as security monitoring, partial face recognition is a challenging but important issue, because the captured faces in real-world surveillance videos may be occluded or with variations. Though current face recognition methods perform well in relatively constrained scenes, they may suffer from degradation for partial faces. In this paper, we propose a framework of model-based transfer learning and sparse coding (MTLSC) for partial face recognition. First, due to less information in partial face image, we exploit the mirrored image of an original probe sample as sample augment to provide further information. Considering the inadequacy of training face samples, we obtain face features based on model-based transfer learning VGGNet that is pre-trained on VGGFace dataset. Then we reconstruct face features by sliding window in view of different sizes of partial face hard to extract the same feature dimension. Finally we carry out sparse coding with rectification and calculate the minimum score of the probe and mirrored samples among all classes to get the results. Thus, by model-based transfer learning, sliding window for feature reconstruction and sparse coding with rectification, the proposed framework improves partial face recognition performance. Experimental results on three face databases (LFW, AR and NIR), and two person re-identification databases (iLIDS-VID and PKU-Reid) demonstrate our method is effective for partial face recognition.

In this paper, we propose the Glyph-Semanteme fusion Embedding (GSE) for Chinese character and apply it to Offline Handwritten Chinese Text Recognition (offline-HCTR). It is well known that the number of Chinese characters is very large and the glyphs of these characters are complex, but few researchers realize that the underlying reason for this phenomenon is that Chinese is a form of ideogram, which indicates that there are correlations between the glyph and semanteme of a character. In order to utilize this feature and create better representations for Chinese characters, firstly, we extract the glyph embedding and semanteme embedding for each Chinese character; then we propose a parameterized gated fusion strategy to automatically calculate the Glyph-Semanteme fusion Embedding for each character by fusing its glyph embedding and semanteme embedding. We apply the proposed GSE to an attention-based Encoder-decoder network for the offline-HCTR task. Furthermore, two kinds of GSE, Character-level GSE (CGSE) and Text-level GSE (TGSE), are applied to the decoder phase to yield the predictions. On the standard benchmark ICDAR-2013 HCTR competition dataset, the proposed method achieves 96.65% character-level recognition accuracy, which demonstrates the effectiveness of the proposed glyph-semanteme fusion embedding.

Document image classification plays a vital role in the document image processing system. Thus it is of great importance to have a clear understanding of the state-of-the-art of the document image classification field, especially in this deep learning era, which will facilitate the development of effective document image processing systems. In this paper, we provide a comprehensive survey of the progress that has been made in the field of document image classification over the past two decades. We categorize the document images into non-mobile images and mobile images according to the way they are acquired. The existing document image classification methods for these two types of images are reviewed, which are classified as textual-based methods, structural-based methods, visual-based methods and hybrid methods. We further compare the performance of different classification methods on several public benchmark datasets. Finally, we highlight some open issues and recommend promising directions for future research.

A local optimal matching scheme of attributed scattering centres (ASC) is proposed for occluded target recognition in synthetic aperture radar (SAR) images. Considering the number of matching pairs...

The state-of-the-art semantic segmentation solutions usually leverage different receptive fields via multiple parallel branches to handle objects of different sizes. However, employing separate kernels for individual branches may degrade the generalization of the network to objects with different scales, and the computational cost increases with the increase of the number of branches. To tackle this problem, we propose a novel network structure, namely Kernel-Sharing Atrous Convolution (KSAC), where branches with different receptive fields share the same kernel, i.e., let a single kernel ‘see’ the input feature maps more than once with different receptive fields. Experiments conducted on the benchmark PASCAL VOC 2012 dataset show that our proposed sharing strategy can not only boost the network’s generalization and representation abilities but also reduce the computational cost significantly. Specifically, on the validation set, when compared with DeepLabv3+, about 2.7G FLOPs and 12.7G FLOPs are saved for output stride = 16 and 8 respectively. In addition, different from the widely used ASPP structure, our proposed KSAC is able to further improve the mIOU by taking benefit of wider context with larger atrous rates. Finally, our KSAC achieves mIOUs of 88.1%, 45.47% and 80.7% on the PASCAL VOC 2012 test set (Everingham et al., 2009), ADE20K dataset (Zhou et al., 2017) and Cityscapes datasets (Marius et al., 2016), respectively. Our full code will be released on Github: https://github.com/edwardyehuang/iSeg.

This paper addresses the problem of consensus-based state estimation for interval type-2 Takagi–Sugeno (T–S) fuzzy time-delay systems over a wireless sensor network under deception attacks, partial information exchange and switching sensor-network topologies. The switching-topology-dependent estimators are constructed using interval type-2 fuzzy models considering the cases where the premise variables are available and unavailable. By means of the mode-dependent dwell time approach, sufficient conditions are established such that the estimation error system is exponentially and ultimately bounded. In addition, a fuzzy Lyapunov functional is utilized to derive the existence conditions of estimators for the case that the premise variable are unavailable. Two illustrative examples are provided to verify the effectiveness of the proposed method.

In recent years, person re-identification based on video has become a hot topic in the field of person re-identification. The self-attention mechanism can improve the ability of deep neural networks in computer vision tasks such as image classification, image segmentation and natural language processing tasks. In order to verify whether the self-attention can improve the performance or not in person re-identification tasks, this paper applies two self-attention mechanisms, non-local attention and recurrent criss-cross attention to person re-identification model, and experiments are conducted on Market-1501, DukeMTMC-reID and MSMT17 person re-identification datasets. The results show that the self-attention mechanism can improve the accuracy of the person re-identification model. The accuracy is higher when the self-attention module is inserted into the convolutional layers of the re-identification network.

Recent years have witnessed the emerging development of flexible electronic devices in applications, such as human gait monitoring. The automatic posture identification could be achieved by sensing the feet pressure. However, the sensor data processing chip usually consumes high power. In this letter, a wearable gait recognition system integrated with an ultrahigh-energy efficiency chip is proposed. A low-power human gait pressure monitoring and recognition solution that uses artificial intelligence near-sensor chips is reported. A large-scale flexible pressure sensor array (64 × 64 pixels) was fabricated. The gait posture was recognized dynamically with a CMOS chip embedded near the sensors. The system owns a recognition rate for three different gaits of 92% with the chip power consumption of 1.45 mW. This letter paves the way for future low-power human gait monitoring.

There is a gap between recent development of 4K display technologies and the short storage of 4K contents. Super-Resolution (SR) serves as a bridge to harmonize the need and demand. Recently, Convolutional Neural Network (CNN) based networks have demonstrated great property in image SR. However, most existing methods require large model capacity and consume expensive computation for high performance. Besides, most methods keep the upscaling part relatively simple compared with the feature extraction part. For feature fusion, some methods directly concatenate the features of multi-levels, which is suboptimal due to ignoring the importance of different features. In this work, we propose a recursive multi-stage upscaling network (RMUN) with multiple sub-upscaling modules (SUMs) and a discriminative self-ensemble module (SEM). Specifically, we extract local hierarchical features by using a novel feature extraction module (FEM) which is recursive to reduce the number of parameters. Then, we construct multiple sub-upscaling modules to produce various high-resolution features in forward propagation. This strategy enhances the upscaling part and provides multiple error feedback routes. Furthermore, we employ an SEM for global hierarchical feature recalibration, which can selectively emphasize informative features and surpass less useful ones. Extensive quantitative and qualitative evaluations on benchmark datasets show that our proposed method performs comparable with the state-of-the-art methods in terms of the balance of model size and model performance.

Nonrigid image registration plays an important role in the field of computer vision and medical application. The methods based on Demons algorithm for image registration usually use intensity difference as similarity criteria. However, intensity based methods can not preserve image texture details well and are limited by local minima. In order to solve these problems, we propose a Gabor feature based LogDemons registration method in this article, called GFDemons. We extract Gabor features of the registered images to construct feature similarity metric since Gabor filters are suitable to extract image texture information. Furthermore, because of the weak gradients in some image regions, the update fields are too small to transform the moving image to the fixed image correctly. In order to compensate this deficiency, we propose an inertial constraint strategy based on GFDemons, named IGFDemons, using the previous update fields to provide guided information for the current update field. The inertial constraint strategy can further improve the performance of the proposed method in terms of accuracy and convergence. We conduct experiments on three different types of images and the results demonstrate that the proposed methods achieve better performance than some popular methods.

Although a substantial number of traffic videos have been accumulated via daily monitoring, deep learning is seldom utilized to process these data for multilevel traffic state detection. The application of deep learning is limited for two reasons: (1) the multilevel traffic state based on traffic images has not been defined. (2) The high noise information in traffic images and extremely similar features of adjacent traffic states hinder accurate detection. Based on this situation, A new definition of the image-based multilevel traffic state is proposed using the ratio of the vehicle areas to the road areas in a traffic image, and a standard image dataset, including various illuminations and vast scenes, are established. A deep residual network named TrafficNet, which is embedded with Squeeze-and-Excitation blocks and is learned by the improved triplet loss, is proposed for multilevel traffic state detection. The Squeeze-and-Excitation block effectively reduces the model's attention to noise information and focuses on road areas that are associated with traffic features in an image. The improved triplet loss maps the learned features to a metric space where the distance between features of inter-class is larger than that within the same class, which improves the discrimination of features between adjacent traffic states. Relevant experiments prove that the performance of TrafficNet, whose accuracy (Acc) in classifying 10 traffic states reaches 94.27% with the testing dataset, is much better than that of traditional deep classification models, which do not include Squeeze-and-Excitation blocks or the improved triplet loss.

This paper investigates a fuzzy intercontinental multimodal routing problem with uncertainties in time and train capacity. The transport network has characteristics of a long distance across continents and a road-rail multimodal routing, with four kinds of nodes and three kinds of arcs. Based on a variant of the vehicle routing problem, the tractor and semi-trailer routing problem is considered for the freight collection part for intercontinental trains. Additionally, the rail routing problem includes domestic direct trains, and intercontinental trains with hard time windows of departure. To make this problem more applicable to real-world circumstances, we describe two types of uncertainty parameters, including time and train uncertainties. Based on the transport conditions of stations, the time uncertainty is considered. Due to the multimodal transport stations' operating capacity and container collection methods, train capacity uncertainty is taken into account. Furthermore, we use solution methods based on the defuzzification approach to solve a fuzzy mixed integer linear programming model and generate a series of instances to verify the fuzzy model. We perform sensitivity analyses of the parameters. The results show that different quantities of intercontinental trains can change the performance by 10% to 20%. The objective function may decrease by more than 20% when the service level increases by 0.1. A sensitivity analysis of the time satisfaction confidence level also shows the trends of fuzzy time and the objective function. These analyses can give reference results about timeliness, transport resource allocation and other suggestions for the intercontinental multimodal transport routing problem.

This paper aims to improve users’ experience in affective interactive installations through the diversification of interfaces. With logically organized hierarchical experience, diverse interfaces with emotion data as inputs enhance users’ emotional interaction to be more natural and immersive. By using facial affect detection technology, an installation with diverse input interfaces was tested with an organic formal setting. Mechanical flowers and support structure based on the organic form were deployed as its physical output for a multitude of sensorial dimensions. With actions of the mechanical flowers, such as blooming, closing, rotating, glowing and blinking, a layered experiential sequence was created and the atmosphere of the installation was evaluated to be more engaging. In this way, the layered complexity of information was transferred to users’ immersive emotional experience. We believe that the practices in this work can contribute to deeper emotional engagement with users and add new layers of emotional interactivity.

With the rapid development of deep neural networks, underwater vision plays an increasingly important role in the underwater robotic operation. However, the scarce underwater datasets greatly limit the performance of deep learning on underwater visual tasks, further hindering the applications of underwater operation. To solve this problem, we propose an underwater image synthesis method, which can directly convert the natural light image into the synthetic underwater image end-to-end. Particularly, a pixel-level self-supervised training strategy is designed to maximize the structural similarity between the synthesized and real images, through training the real underwater images. Finally, extensive experiments are carried out, and the obtained results demonstrate the effectiveness and superiority of our methods by quantitative and qualitative comparisons. The proposed underwater image synthesis method offers a valuable sight for underwater vision and manipulating control.

In a variety of fields, in particular those involving imaging and optics, we often measure signals whose phase is missing or has been irremediably distorted. Phase retrieval attempts to recover the phase information of a signal from the magnitude of its Fourier transform to enable the reconstruction of the original signal. Solving the phase retrieval problem is equivalent to recovering a signal from its auto-correlation function. In this paper, we assume the original signal to be sparse; this is a natural assumption in many applications, such as X-ray crystallography, speckle imaging and blind channel estimation. We propose an algorithm that resolves the phase retrieval problem in three stages, first, we leverage the finite rate of innovation sampling theory to super-resolve the auto-correlation function from a limited number of samples, second, we design a greedy algorithm that identifies the locations of a sparse solution given the super-resolved auto-correlation function, finally, we recover the amplitudes of the atoms given their locations and the measured auto-correlation function. Unlike traditional approaches that recover a discrete approximation of the underlying signal, our algorithm estimates the signal on a continuous domain, which makes it the first of its kind. Along with the algorithm, we derive its performance bound with a theoretical analysis and propose a set of enhancements to improve its computational complexity and noise resilience. Finally, we demonstrate the benefits of the proposed method via a comparison against Charge Flipping, a notable algorithm in crystallography.

Distributed arithmetic (DA) brings area and power benefits to digital designs relevant to the Internet-of-Things. Therefore, a new error resilient technique for DA computation is proposed to improve robustness against process, voltage, and temperature variations. The proposed approach mitigates the effect of timing violations by first providing a guardband for significant (most significant bit) computations. This guardband is initially achieved by modifying the order of DA serial operations and borrowing time from the least significant bit (LSB) group. Therefore, LSB computation can correspond to the critical path, and timing error can be tolerated at the cost of acceptable accuracy loss. Moreover, the shifted-phase clock signals are applied on the end-point registers, thereby increasing the global guardband without any effect on system sampling rate. Our approach is demonstrated on a 16-tap FIR filter using the 65 nm CMOS process. The simulation results demonstrate that this design can maintain error-free operation without worst case timing margin, and achieve up to 42% power savings by voltage scaling when the worst case margin is considered. This is at the cost of a 6.3% delay and 7.3% overhead.

This paper presents an experimental, numerical, and analytical study on the crushing of web girders, to provide insights into the behaviour of a ship’s structure in a collision or grounding inciden...

Intra-pulse modulation phase calibration is necessary in inverse synthetic aperture radar (ISAR) imaging of high-speed targets. Traditional intra-pulse phase error compensation strategies rarely handle the high-order and slow-time-variant phase components induced during the coherent processing interval. In this paper, a novel intra-pulse modulation phase calibration with a two-dimensional (2-D) parametric phase model is proposed. It models the intra-pulse phase errors as a 2-D time-variant polynomial with accommodation of both fast-time and slow-time modulation. Entropy minimization of high-resolution range profiles (HRRPs) is developed to retrieve the phase error parameters. Improved coordinate descent optimization solver is established by Levenberg-Marquardt (LM) algorithm in order to find the global optimum of entropy efficiently. Comparative experiments using both simulated and real measured data are performed to demonstrate the enhancements of the proposed algorithm.

Font generation is a challenging problem especially for some writing systems that consist of a large number of characters and has attracted a lot of attention in recent years. However, existing methods for font generation are often in supervised learning. They require a large number of paired data, which is labor-intensive and expensive to collect. Besides, common image-to-image translation models often define style as the set of textures and colors, which cannot be directly applied to font generation. To address these problems, we propose novel deformable generative networks for unsupervised font generation (DGFont). We introduce a feature deformation skip connection (FDSC) which predicts pairs of displacement maps and employs the predicted maps to apply deformable convolution to the low-level feature maps from the content encoder. The outputs of FDSC are fed into a mixer to generate the final results. Taking advantage of FDSC, the mixer outputs a high-quality character with a complete structure. To further improve the quality of generated images, we use three deformable convolution layers in the content encoder to learn style-invariant feature representations. Experiments demonstrate that our model generates characters in higher quality than state-of-art methods. The source code is available at https://github.com/ecnuycxie/DG-Font., Accepted by CVPR-2021

Paper influence analysis is essential technology of the literature management system. In recent years, research on influence analysis has received crucial attention from the industry and academia. In this field, there are many new challenges in how to evaluate valuable papers in the paper citation network. Therefore, in this paper, we have proposed an influence model in paper citation networks with integration features of PageRank and Hyperlink-Induced Topic Search (HITS). First, three ranking features are extracted with the combination of PageRank value, hub, and authority. Second, more features are added to our ranking classification method. Third, the ranking-based classification with $P$ ageRank and HITS is proposed to analyze the paper propagation. An influence score is computed to evaluate the importance of a paper. The experiments show that our ranking-based influence model is effective for paper propagation.

BACKGROUND: With the rapid development of deep learning, several neural network models have been proposed for automatic segmentation of prohibited items. These methods usually based on a substantial amount of labelled training data. However, for some prohibited items of rarely appearing, it is difficult to obtain enough labelled samples. Furthermore, the category of prohibited items varies in different scenarios and security levels, and new items may appear from time to time. OBJECTIVE: In order to predict prohibited items with only a few annotated samples and inspect prohibited items of new categories without the requirement of retraining, we introduce an Attention-Based Graph Matching Network. METHODS: This model applies a few-shot semantic segmentation network to address the issue of prohibited item inspection. First, a pair of graphs are modelled between a query image and several support images. Then, after the pair of graphs are entered into two Graph Attention Units with similarity weights and equal weights, the attentive matching results will be obtained. According to the matching results, the prohibited items can be segmented from the query image. RESULTS: Experiment results and comparison using the Xray-PI dataset and SIXray dataset show that our model outperforms several other state-of-the-art learning models. CONCLUSIONS: This study demonstrates that the similarity loss function and the space restriction module proposed by our model can effectively remove noise and supplement spatial information, which makes the segmentation of the prohibited items on X-ray images more accurate.

Few-shot image segmentation intends to segment query images (test images) given only a few support samples with annotations. However, previous works ignore the impact of the object scales, especially in the support images. Meanwhile, current models only work on images with the similar size of the object and rarely test on other domains. This paper proposes a new few-shot segmentation model named DCNet, which fully exploits the support set images and their annotations and is able to generalize to the test images with unseen objects of various scales. The idea is to gradually compare the features from the query and the support image, and refine the features for the query. Furthermore, a sequential k-shot comparison method is proposed based on the ConvGRU to integrate features from multiple annotated support images. Experiments on Pascal VOC dataset and X-ray Security Images demonstrate the excellent generalization performance of our model.

We use a comprehensive set of performance metrics to analyze the improvement in the classification power and prediction accuracy of various bankruptcy prediction models after adding governance variables and/or varying the estimation method used. In a sample covering bankruptcies of U.S. public firms in the period 2000 to 2015, we find that the addition of governance variables significantly improves the performance of all bankruptcy prediction models. We also find that the additional explanatory power provided by governance measures improves the further the firm is from bankruptcy, which suggests that governance variables may provide earlier and more accurate warning of the firm's bankruptcy potential. Our findings show that the performance of any bankruptcy prediction model is significantly affected by the estimation method used. We find that regardless of the bankruptcy model, hazard analysis provides the best classification and out-of-sample forecast accuracy among the parametric methods. Furthermore, non-parametric methods such as neural networks, data envelopment analysis or classification and regression trees appear to provide comparable and sometimes superior classification accuracy to hazard analysis. Lastly, we use the dynamic panel generalized methods of moments model to address concerns raised in prior studies about the susceptibility of similar studies to endogeneity issues and find that our findings continue to hold.

Range alignment is an essential procedure in the translation motion compensation of inverse synthetic aperture radar imaging. Global optimization or maximum-correlation-based algorithms have been used to realize range alignment. However, it is still challenging to achieve range alignment in low signal-to-noise ratio scenarios, which are common in inverse synthetic aperture radar imaging. In this paper, a novelanti-noise range alignment approach is proposed. In this new method, the target motion is modelled as a uniformly accelerated motion during a short sub-aperture time. Minimum entropy optimization is implemented to estimate the motion parameters in each sub-aperture. These estimated parameters can be used to align the profiles of the current sub-aperture. Once the range profiles of eachsub-aperture are aligned, the non-coherent accumulation gain is obtained by averaging all profiles in each sub-aperture, which can be used as valuable information. The accumulation and correlation method is applied to align the average range profiles of each sub-aperture because the former step focuses mainly on alignment within the sub-apertures. Experimental results based on simulated and real measured data demonstrate the effectiveness of the proposed algorithm in low signal-to-noise ratio scenarios.

Transfer learning seeks to improve the generalization performance of a target task by exploiting the knowledge learned from a related source task. Central questions include deciding what information one should transfer and when transfer can be beneficial. The latter question is related to the so-called negative transfer phenomenon, where the transferred source information actually reduces the generalization performance of the target task. This happens when the two tasks are sufficiently dissimilar. In this paper, we present a theoretical analysis of transfer learning by studying a pair of related perceptron learning tasks. Despite the simplicity of our model, it reproduces several key phenomena observed in practice. Specifically, our asymptotic analysis reveals a phase transition from negative transfer to positive transfer as the similarity of the two tasks moves past a well-defined threshold.

In this paper, we analyze the relationship between the performance of the text-independent feature and text factors of the handwriting on Chinese writer identification. Text factors contain two types of information: the number of characters in both query and reference and the number of the same characters in both query and reference. We conclude that the performance increases when the query and reference contain more characters, and the minimum number of needed characters is 50. The number of the same characters in both query and reference has little influence on the identification when the number of characters is more than 50. The conclusions are verified by repeated writer identification tests with different amount of characters on the handwriting document pages.

Scene text transfer for cross-language aims to erase the original scene text and generate another language text image into the original scene text image with the same style, including the style of fonts, colors, size, and background texture. Scene text transfer for cross-language is a challenging problem as the complicated background scene and a huge difference between languages, which demanding high-quality performance for both text transfer and text erasing. In this work, we propose a scene text transfer framework for cross-language which consists of three steps: regional text extraction, style transfer, and scene text combination. The regional text extraction is designed to crop the text region of a natural scene image and transform it to be a rectangle text image. In the second step, a style transfer network is proposed to retain the style of text image and transfer the text content. In the step of the scene text combination, our model combines the rendered text image with the original scene image to produce the final result. In the optimization part, we introduce a novel background consistent loss to improve the performance of background generation. Experiments demonstrate that our framework generates scene text images of higher quality than previous methods.

In the practical application of medical image analysis, due to the different data distributions of source domain and target domain and the lack of the labels of target domain, domain adaptation for unsupervised cross-domain classification attracts widespread attention. However, current methods take knowledge transfer model and classification model as two separate training stages, which inadequately considers and utilizes the intrinsic information interaction between modules. In this paper, we propose a coherent cooperative learning framework based on transfer learning for unsupervised cross-domain classification. The proposed framework is constructed by two classifiers trained by transfer learning, which can respectively classify images of source domain and target domain, and a Wasserstein CycleGAN for image translation and data augmentation. In the coherent process, all modules are updated in turn, and the data is transferred between different modules to realize the knowledge transfer and collaborative training. The final prediction is obtained by a voting result of two classifiers. Experimental results on three pneumonia databases demonstrate the effectiveness of our framework with diverse backbones.

Similarity analysis of DNA sequences can clarify the homology between sequences and predict the structure of, and relationship between, them. At the same time, the frequent patterns of biological sequences explain not only the genetic characteristics of the organism, but they also serve as relevant markers for certain events of biological sequences. However, most of the aforementioned biological sequence similarity analysis methods are targeted at the entire sequential pattern, which ignores the missing gene fragment that may induce potential disease. The similarity analysis of such sequences containing a missing gene item is a blank. Consequently, some sequences with missing bases are ignored or not effectively analyzed. Thus, this paper presents a new method for DNA sequence similarity analysis. Using this method, we first mined not only positive sequential patterns, but also sequential patterns that were missing some of the base terms (collectively referred to as negative sequential patterns). Subsequently, we used these frequent patterns for similarity analysis on a two-dimensional plane. Several experiments were conducted in order to verify the effectiveness of this algorithm. The experimental results demonstrated that the algorithm can obtain various results through the selection of frequent sequential patterns and that accuracy and time efficiency was improved.

The public opinion monitoring system can monitor the latest information at any time and manage the network information activities by focusing on the website information about hot issues and key areas. The traditional Lambda architecture does not write abstract data on the storage layer, and the data latency and cost are relatively high on the storage. In this paper, we use the improved Lambda-based architecture to support stream and batch computing to make computational logic reused. The improved Lambda architecture enables the storage of unified historical data in full volume and online real-time incremental data. The storage can support high-throughput real-time write, as well as multi-field combination search and full-text retrieval.

There are many artistic support systems available for artists, but none of them provides high-grade collaborative support for traditional artistic painting as an assistant. To collaborate with human artists, the support system should involve essential artistic procedures and be capable of imitating basic artistic techniques of human artists. In this paper, we address this issue and introduce a novel painting support system, namely Skywork-daVinci, based on a human-in-the-loop mechanism. A CPSS-based approach is proposed for collaboration between human artists and machines. Based on the approach, we develop a robotic painting system to imitate painting techniques of artists. Then, an acrylic painting case is presented to demonstrate the feasibility of our system. Furthermore, comparisons between existing robotic painting systems and ours are discussed. As experimental results show, our robotic painting system could imitate the essential painting techniques of human artists and reproduce the paintings drawn by them. This primary system acts as the basis to realize our collaborative artistic support system entirely in future work.

Emotion recognition based on electroencephalography (EEG) signals is a current focus in brain-computer interface research. However, the classification of EEG is difficult owing to large amounts of data and high levels of noise. Therefore, it is important to determine how to effectively extract features that include important information. Regularization, one of the effective methods for EEG signal processing, can effectively extract important features from the signal and has potential applications in EEG emotion recognition. Currently, the most popular regularization technique is Lasso (L1) and Ridge Regression (L2). In recent years, researchers have proposed many other regularization terms. In theory, Lq-type regularization has a lower q value, which means that it can be used to find solutions with better sparsity. L1/2 regularization is of Lq type (0 < q < 1) and has been shown to have many attractive properties. In this work, we studied the L1/2 penalty in sparse logistic regression for three-classification EEG emotion recognition, and used a coordinate descent algorithm and a univariate semi-threshold operator to implement L1/2 penalty logistic regression. The experimental results on simulation and real data demonstrate that our proposed method is better than other existing regularization methods. Sparse logistic regression with L1/2 penalty achieves higher classification accuracy than the conventional L1, Ridge Regression, and Elastic Net regularization methods, using fewer but more informative EEG signals. This is very important for high-dimensional small-sample EEG data and can help researchers to reduce computational complexity and improve computational accuracy. Therefore, we propose that sparse logistic regression with the L1/2 penalty is an effective technique for emotion recognition in practical classification problems.

Confidential information firewalling with text classifiers is to recognize the text containing confidential information whose publication might pose a threat to national security, business trade, or personal life. Word embedding is a component of the detector and plays an important role. Existing word embeddings, e.g., Word2Vec, fail to learn a clear task classification boundary, i.e., the confidential polarities of words are opposite but the embedding vectors of the words are close to each other. We propose a confidentiality-oriented word embedding, CES2Vec, for confidential information detection. We embed confidentiality into semantics to catch both of them together, which can learn the word embedding with a clear task classification boundary. We use real-world data from WikiLeaks and conduct the comparison experiments of our CES2Vec and popular methods. The experimental results show that our proposed method is better than the previously reported methods in detecting confidential information.

Malware has become a serious threat to network security. Traditional static analysis methods usually cannot effectively detect packers, obfuscations, and variants. Dynamic analysis is not efficient when dealing with large amounts of malware. Aiming at the shortcomings of the existing methods, this paper proposes a method for analyzing malware based on the capsule network. It uses a supervised learning method to train the capsule network with a large number of malware samples with existing category labels. In the process of constructing features, this paper adopts a method of combining static features and dynamic features to extract the operation code information based on static analysis, and extract the API call sequence information based on general analysis. Both characteristics can well represent the structure and behavior of malware. Then use N-Gram to construct sequence features, visualize the N-Gram sequence, generate malware images, and finally use the capsule network for classification detection. In addition, this paper improves the original capsule network and verifies the effect of the improved model.

Objects detection can be regard as the segmentation of foreground from background. In this paper, we propose a foreground segmentation method based on sparse representation of direction features for threat object detection in X-ray images. The threat objects are supposed as foreground and all other contents in the images are background. We extract the direction features to make up foreground dictionary firstly. Then we search the foreground area in the test image through sparse representation of their direction features by foreground dictionary. The experimental results show that this proposed method is robust to the X-ray images with different backgrounds.

Confidential information analysis can identify the text containing confidential information, thereby protecting organizations from the threat posed by leakage of confidential information. It is effective to build a confidential information analyzer based on a neural network. Most of the existing studies pursue high accuracy to design complex networks, ignoring speed and consumption. The optimal defense is to automatically analyze confidential information without compromising routine services. In this paper, we introduce a lightweight network, DSCNN, that can be adapted to low-resource devices. We also introduce two hyper-parameters to balance accuracy and speed. Our motivation is to simplify convolutions by breaking them down because the space dimension and channel dimension are not closely related in the convolutions. Experimental results on real-world data from WikiLeaks show that our proposed DSCNN performs well for confidential information analysis.

The quality control of LCD manufacturing process is very important to minimize cost and maximize product quality. This paper designs a light convolution neural network (CNN) model with fewer parameters that detects surface defects and identifies their types in thin film transistor liquid crystal display (TFT-LCD). We choose different sizes of small convolution kernels to extract the shallow features of the image, and introduce the sparse convolution structure to extract the multi-scale deep features. In addition, because there are few defects in actual industrial production lines, the defect sample numbers may be not sufficient for obtaining much more local features directly based on small sample learning. We use the deep convolutional generative adversarial network (DCGAN) to generate data. The network parameters of the designed model are updated by the re-learning the generated samples based on the original sample and DCGAN. The original small samples are directly sent to the deep feature extraction layer to further strengthen training, so that the model has the ability of continue learning. We make a series of experiments on the real images using the trained model combined with a sliding window technique to detect and classify the detects in the original images. The experimental results show that the model can effectively improve the detection rate and reduce the missed detection rate.

Sarcasm detection is to identify the text with the author’s sarcastic attitude. Verbal sarcasm is one main error sources of sentiment analysis tasks. However, labeled sarcastic samples are expensive to obtain. Previous approaches, e.g., model user and topic embedding from multiple perspectives together with large-scale network training, are not suitable for real business scenarios that expect low cost and high speed. In this paper, we propose a semi-supervised method for contextual sarcasm detection in online discussion forums. We adopt author and topic sarcastic prior preference as context embedding that supply simple but representative background knowledge. Then we introduce a sarcasm-unlabeled learning method to utilize a few labeled sarcastic samples and model the classification boundary. Experiments are conducted on real-world data from Reddit, and the results indicate the outperformance over existing methods.

For intelligent four-wheel-drive (4WD) electric vehicle (EV), the vehicle speed can be planned and controlled for energy saving based on the slope information of road ahead. To reduce the calculation load of the optimisation algorithm, the model predictive control (MPC) method is formulated based on the time horizon in this study. Furthermore, a fast gradient method based control tool-GARMPC is used to solve the optimisation problem. First, the longitudinal dynamics model of 4WD EV based on time horizon and distance horizon is established based on the road slope information, respectively. Second, the MPC problem based on the time-discrete model is formulated and solved by GARMPC tool. For comparison, a dynamic program (DP) control method is introduced based on the distance-discrete model. Finally, the simulation is conducted under a designed road condition and a real measured road condition. The results show that the time-horizon based MPC method can significantly reduce the energy consumption compared with the proportion integration differentiation control method, which is similar to the driver's operation. Compared with the DP optimisation method, the time-based MPC method reduces the calculation time to smaller than 1 ms, which is essential for real-time application in a road vehicle.

For many modern applications in science and engineering, data are collected in a streaming fashion carrying time-varying information, and practitioners need to process them with a limited amount of memory and computational resources in a timely manner for decision making. This often is coupled with the missing data problem, such that only a small fraction of data attributes are observed. These complications impose significant, and unconventional, constraints on the problem of streaming Principal Component Analysis (PCA) and subspace tracking, which is an essential building block for many inference tasks in signal processing and machine learning. This survey article reviews a variety of classical and recent algorithms for solving this problem with low computational and memory complexities, particularly those applicable in the big data regime with missing data. We illustrate that streaming PCA and subspace tracking algorithms can be understood through algebraic and geometric perspectives, and they need to be adjusted carefully to handle missing data. Both asymptotic and non-asymptotic convergence guarantees are reviewed. Finally, we benchmark the performance of several competitive algorithms in the presence of missing data for both well-conditioned and ill-conditioned systems., 27 pages, 7 figures, submitted to the Proceedings of IEEE

The modified quadratic discriminant function (MQDF) is an effective classifier for handwritten Chinese character recognition (HCCR). However, it suffers from high memory requirement for the storage of its parameters, which makes it impractical to be embedded in memory limited hand-held devices. In this paper, we explore the applicability of sparse coding to build compact MQDF classifiers. To be specific, we use sparse coding to compact the parameters of MQDF. Two methods of sparse coding, viz., the maximum likelihood-based method and the K-SVD method, are adopted to build two compact MQDF classifiers, namely, MQDF-ML classifier and MQDF-KSVD classifier. Furthermore, we learn multiple dictionaries rather than single dictionary for sparse coding, because the multiple dictionary learning is capable of not only greatly reducing the computational complexity, but also alleviating the degradation of recognition accuracy, compared to the single dictionary learning. Experiments and comparison with the existing method have demonstrated the effectiveness of our proposed method for the issue of unconstrained handwritten Chinese character recognition.

A powerful platform of digital brain is proposed using crowd wisdom for brain research, based on the computational artificial intelligence model of synthesis reasoning and multi-source analogical generating. The design of the platform aims to make it a comprehensive brain database, a brain phantom generator, a brain knowledge base, and an intelligent assistant for research on neurological and psychiatric diseases and brain development. Using big data, crowd wisdom, and high performance computers may significantly enhance the capability of the platform. Preliminary achievements along this track are reported.