Search

Digitized handwritten documents have been used for various purposes, including age detection, a crucial area of research in fields like forensic investigation and medical diagnosis. Automated age recognition is deemed to be a difficult task, due to the great degree of similarity and overlap across people's handwriting. Consequently, the efficiency of the classification system is determined by extracting pertinent features from handwritten documents. This research proposes a set of age-related features suggested by a graphologist to classify handwritten documents into two age groups: youth adult and mature adult. The extracted features are: slant irregularity (SI), pen pressure irregularity (PPI), text lines irregularity (TLI) and the percentage of black and white pixels (PWB). Support Vector Machines (SVM) and Neural Network (NN) classifiers have been used to train, validate and test the proposed approach using two different datasets: the FSHS and the Khatt datasets. When applied to the FSHS dtaset using SVM and NN approaches, the proposed method resulted in a classification accuracy of 71% and 63.5%, respectively. Meanwhile, when applied to the Khatt dataset, our method outperformed state-of-the-art methods with a classification accuracy of 65.2% and 67% utilising SVM and NN classifiers, respectively. These are the best rates available right now in this field. [ABSTRACT FROM AUTHOR]

• New model for scene text VQA that jointly reasons about textual and visual modalities. • Attending on multi-modal features is better than attending separately to each modality. • Grid features from an object detection backbone proves to work well in this task. • The proposed model is up to 5 times faster than current state of the art models. • Combined with a standard VQA model reaches state of the art on mixed datasets. This paper presents a new model for the task of scene text visual question answering. In this task questions about a given image can only be answered by reading and understanding scene text. Current state of the art models for this task make use of a dual attention mechanism in which one attention module attends to visual features while the other attends to textual features. A possible issue with this is that it makes difficult for the model to reason jointly about both modalities. To fix this problem we propose a new model that is based on an single attention mechanism that attends to multi-modal features conditioned to the question. The output weights of this attention module over a grid of multi-modal spatial features are interpreted as the probability that a certain spatial location of the image contains the answer text to the given question. Our experiments demonstrate competitive performance in two standard datasets with a model that is × 5 faster than previous methods at inference time. Furthermore, we also provide a novel analysis of the ST-VQA dataset based on a human performance study. Supplementary material, code, and data is made available through this link. [ABSTRACT FROM AUTHOR]

• A mental stress state diagnostic for timely intervention from clinicians. • Deep neural network with hierarchical learning capabilities using IoT based biomarkers. • Learns using both wrist-based and chest-based sensor bio-signal features. • Model-level fusion of high-level representations for stress state classification. Affective state recognition at an early stage can help in mood stabilization, stress and depression management for mental well-being. Pro-active and remote mental healthcare warrants the use of various biomarkers to detect the affective mental state of the individual by evaluating the daily activities. With the easy accessibility of IoT-based sensors for healthcare, observable and quantifiable characteristics of our body, physiological changes in the body can be measured and tracked using various wearable devices. This work puts forward a model for mental stress state detection using sensor-based bio-signals. A multi-level deep neural network with hierarchical learning capabilities of convolution neural network is proposed. Multivariate time-series data consisting of both wrist-based and chest-based sensor bio-signals is trained using a hierarchy of networks to generate high-level features for each bio-signal feature. A model-level fusion strategy is proposed to combine the high-level features into one unified representation and classify the stress states into three categories as baseline, stress and amusement. The model is evaluated on the WESAD benchmark dataset for mental health and compares favourably to state-of-the-art approaches giving a superlative performance accuracy of 87.7%. [ABSTRACT FROM AUTHOR]

• Enhance face verification performance for demographics effected by decision bias. • Operates on comparison score-level to enable using existing face recognition models. • Based on the notation of individual fairness to treat similar face groups similarly. • Effectiveness proved on three databases and two face embeddings. • Unsupervised enhancement of the overall and intra-class recognition performance. Current face recognition systems achieve high progress on several benchmark tests. Despite this progress, recent works showed that these systems are strongly biased against demographic sub-groups. Consequently, an easily integrable solution is needed to reduce the discriminatory effect of these biased systems. Previous work mainly focused on learning less biased face representations, which comes at the cost of a strongly degraded overall recognition performance. In this work, we propose a novel unsupervised fair score normalization approach that is specifically designed to reduce the effect of bias in face recognition and subsequently lead to a significant overall performance boost. Our hypothesis is built on the notation of individual fairness by designing a normalization approach that leads to treating "similar" individuals "similarly". Experiments were conducted on three publicly available datasets captured under controlled and in-the-wild circumstances. Results demonstrate that our solution reduces demographic biases, e.g. by up to 82.7% in the case when gender is considered. Moreover, it mitigates the bias more consistently than existing works. In contrast to previous works, our fair normalization approach enhances the overall performance by up to 53.2% at false match rate of 10 − 3 and up to 82.9% at a false match rate of 10 − 5. Additionally, it is easily integrable into existing recognition systems and not limited to face biometrics. [ABSTRACT FROM AUTHOR]

• We propose a module to predict boundary shapes and boxes with a new masking scheme. • We comprehensively investigate proposed frameworks with various masking methods. • The proposed modularizable component can easily be applied to existing models. • BshapeNet + improves accuracy notably compared to the baselines, Faster and Mask RCNN. • We achieved highly competitive results with SOTA detection and segmentation models. We propose a modularizable component that can predict the boundary shapes and boxes of an image, along with a new masking scheme for improving object detection and instance segmentation. Specifically, we introduce two types of novel masks: a bounding box (bbox) mask and a bounding shape (bshape) mask. For each of these types, we consider two variants—the "Thick" model and the "Scored" model—both of which have the same morphology but differ in ways that make their boundaries thicker. To evaluate our masks, we design extended frameworks by adding a bshape mask (or a bbox mask) branch to a Faster R-CNN, and call this BshapeNet (or BboxNet). Furthermore, we propose BshapeNet+, a network that combines a bshape mask branch with a Mask R-CNN. Among our various models, BshapeNet+ demonstrates the best performance in both tasks. In addition, BshapeNet+ markedly outperforms the baseline models on MS COCO and Cityscapes and achieves highly competitive results with state-of-the-art models. In particular, the experimental results show that our branch works well on small objects and is easily applicable to various models, such as PANet as well as Faster R-CNN and Mask R-CNN. [ABSTRACT FROM AUTHOR]

Highlights • We propose trainable feature extractors for representation of skeleton poses. • We employ the proposed feature extractor for classification of human actions. • We designed an human action classification method based on string kernel. • We carried out experiments on the MHAD, MSRDA and MIVIA-S action datasets. • We publicly released the MIVIA-S dataset for research purpose. Abstract Automatic interpretation of human actions gained strong interest among researchers in patter recognition and computer vision because of its wide range of applications, such as in social and home robotics, elderly people health care, surveillance, among others. In this paper, we propose a method for recognition of human actions by analysis of skeleton poses. The method that we propose is based on novel trainable feature extractors, which can learn the representation of prototype skeleton examples and can be employed to recognize skeleton poses of interest. We combine the proposed feature extractors with an approach for classification of pose sequences based on string kernels. We carried out experiments on three benchmark data sets (MIVIA-S, MSRSDA and MHAD) and the results that we achieved are comparable or higher than the ones obtained by other existing methods. A further important contribution of this work is the MIVIA-S dataset, that we collected and made publicly available. [ABSTRACT FROM AUTHOR]

The aim of this piece of work is to study some topics related to an Hermite interpolation problem on the unit circle. We consider as nodal points the zeros of the para-orthogonal polynomials with respect to a measure in the Baxter class and such that the sequence of the first derivative of the reciprocal of the orthogonal polynomials is uniformly bounded on the unit circle. We study the convergence of the Hermite–Fejér interpolants related to piecewise continuous functions and we describe the sets in which the interpolants uniformly converge to the piecewise continuous function as well as the oscillatory behavior of the interpolants near the discontinuities, where a Gibbs–Wilbraham phenomenon appears. Finally we present some numerical experiments applying the main results and by considering nodal systems of interest in the theory of orthogonal polynomials. [ABSTRACT FROM AUTHOR]

We discuss matrix polynomials expressed in a Newton basis, and the associated polynomial eigenvalue problems. Properties of the generalized ansatz spaces associated with such polynomials are proved directly by utilizing a novel representation of pencils in these spaces. Also, we show how the family of Fiedler pencils can be adapted to matrix polynomials expressed in a Newton basis. These new Newton–Fiedler pencils are shown to be strong linearizations, and some computational aspects related to them are discussed. [ABSTRACT FROM AUTHOR]

Given m + 1 strictly decreasing numbers h 0 , h 1 , … , h m , we give an algorithm to construct a corresponding finite sequence of orthogonal polynomials p 0 , p 1 , … , p m such that p 0 = 1 , p j has degree j and p m − j ( h n ) = ( − 1 ) n p j ( h n ) for all j , n = 0 , 1 , … , m . Using these polynomials, we construct bivariate Lagrange polynomials and cubature formulas for nodes that are points in R 2 where the coordinates are taken from given finite decreasing sequences of the same length and where the indices have the same (or opposite) parity. [ABSTRACT FROM AUTHOR]

Multi-output least-squares support vector regression machines (MLS-SVR) is proposed by Xu et al. [29] to handle multi-output regression problems. However, the prohibitive cost of model selection severely hinders MLS-SVR’s application. In this paper, an efficient gradient-based model selection algorithm for MLS-SVR is proposed. Firstly, a new training algorithm for MLS-SVR is developed, which allows one to obtain the solution vector for each output independently by dealing with matrices of much lower order. Based on the new training algorithm, a new leave-one-out error estimate is derived through virtual leave-one-out cross-validation. The model selection criterion is based on the new leave-one-out error estimate and its derivatives with respect to the hyper-parameters are also derived analytically. Both the model selection criterion and its partial derivatives can be obtained straightway once a training process ended. Finally, the hyper-parameters corresponding to the lowest model selection criterion is obtained through gradient decent method. The effectiveness and generalization performance of the proposed algorithm are validated through experiments on several multi-output datasets. Experiment results show that the proposed algorithm can save computational time dramatically without losing accuracy. [ABSTRACT FROM AUTHOR]

We propose a novel single face image super-resolution method, which is named Face Conditional Generative Adversarial Network (FCGAN), based on boundary equilibrium generative adversarial networks. Without taking any prior facial information, our approach combines the pixel-wise L 1 loss and GAN loss to optimize our super-resolution model and to generate a high-quality face image from a low-resolution one robustly (with upscaling factor 4 × ). Additionally, Compared with existing peer researches, both training and testing phases of FCGAN are end-to-end pipeline without pre/post-processing. To enhance the convergence speed and strengthen feature propagation, the Generator and Discriminator networks are designed with a skip-connection architecture, and both using an auto-encoder structure. Quantitative experiments demonstrate that our model achieves competitive performance compared with the state-of-the-art models based on both visual quality and quantitative criterions. We believe this high-quality face image generated method can impact many applications in face identification and intelligent monitor. [ABSTRACT FROM AUTHOR]

Recently convolutional neural networks (CNN) have demonstrated remarkable performance in various classification problems. In this paper, we also introduce CNN into in-air handwritten Chinese character recognition (IAHCCR) and propose new directional feature maps, named bend directional feature maps. Then we integrate the combination of various types of directional feature maps with the CNN and obtain better recognition performance compared with other methods reported for IAHCCR. For further improving recognition rate, we propose a new data augmentation method dedicated to in-air handwritten Chinese characters. The proposed data augmentation method combines global transformation with local distortion and effectively enlarges the training dataset. Experimental results demonstrate that our proposed methods can greatly improve the recognition rate for IAHCCR. [ABSTRACT FROM AUTHOR]

Decision Trees and Random Forests are among the most popular methods for classification tasks. Two key issues faced by these methods are: how to select the best attribute to associate with a node and how to split the samples given the selected attribute. This paper addresses an important challenge that arises when nominal attributes with a large number of values are present: the computational time required to compute splits of good quality. We present a framework to generate computationally efficient splitting criteria that handle, with theoretical approximation guarantee, multi-valued nominal attributes for classification tasks with a large number of classes. Experiments with a number of datasets suggest that a method derived from our framework is competitive in terms of accuracy and speed with the Twoing criterion, one of few criteria available that is able to handle, with optimality guarantee, nominal attributes with a large number of distinct values. However, this method has the advantage of also efficiently handling datasets with a large number of classes. These experiments also give evidence of the potential of aggregating attributes to improve the classification power of decision trees. [ABSTRACT FROM AUTHOR]

Identification of prescription patterns is a useful and interesting goal from multiple perspectives. The identified prescription patterns may expand the horizons of medical knowledge, and may be evaluated by subject matter experts to label certain patterns as anomalies calling for further investigation; for example, in prescription costs for insurance companies. This paper presents the statistical modeling details of the tag hierarchical topic model (Tag-HTM) and its application to the Health Insurance Review & Assessment Service (HIRA) dataset. The implementation of Tag-HTM revealed a hierarchical structure for medicine symptom distributions, which could constitute a new hierarchical categorization for diseases. The experimental results demonstrate that our generated hierarchical structure can replicate the existing hierarchy, namely ICD-10, which has been created by medical subject matter experts, to a considerable extent. Furthermore, the experiments indicate a quantitative performance improvement; that is, the superior perplexities of Tag-HTM compared to baselines. Moreover, Tag-HTM was able to isolate prescription patterns with higher medical costs as a branch of hierarchical clustering, and this cluster could form a prescription collection of interest to subject matter experts in insurance companies. [ABSTRACT FROM AUTHOR]

In recent years, QR code has been widely used in various society fields, and has provided great convenience in information exchange, commodity payment and website registration. However, the degradation of QR code images can cause difficulty in reading the code information. In this paper, we propose an alternating minimization restoration model for QR code images. Based on the binary prior, we use the binary characteristic and L 0 norm as the regularization terms, and introduce different auxiliary variables to make the complex model become solvable. The bi-level constraint of bar code image maintains the sparseness of image pixels and gradient. The experimental result shows that our algorithm has excellent effect on common blur kernels and multiple noise, and its performance is better than the state-of-the-art algorithms. [ABSTRACT FROM AUTHOR]

Multi set canonical correlation analysis (mCCA), which extends the application of canonical correlation analysis (CCA) to more than two datasets, is a data driven technique that can jointly analyze the relationship amongst multiple (more than two) datasets. However, the conventional mCCA is directly applicable only to multivariate vector data and requires the image data to be reshaped into vectors. This approach fails to consider the spatial structure of the images and in addition, leads to an increase in the computational complexity. In this paper, we propose new two dimensional mCCA algorithms that operate directly on the image data instead of vectorizing them. Face recognition experiments are presented to compare the performances of conventional mCCA and the proposed two dimensional mCCA techniques. Additionally, experiments against fMRI data are conducted to demonstrate the applicability of the proposed approach in multisubject fMRI analysis. [ABSTRACT FROM AUTHOR]

Many modern applications deal with multi-label data, such as functional categorizations of genes, image labeling and text categorization. Classification of such data with a large number of labels and latent dependencies among them is a challenging task, and it becomes even more challenging when the data is received online and in chunks. Many of the current multi-label classification methods require a lot of time and memory, which make them infeasible for practical real-world applications. In this paper, we propose a fast linear label space dimension reduction method that transforms the labels into a reduced encoded space and trains models on the obtained pseudo labels. Additionally, it provides an analytical method to update the decoding matrix which maps the labels into the original space and is used during the test phase. Experimental results show the effectiveness of this approach in terms of running times and the prediction performance over different measures. [ABSTRACT FROM AUTHOR]

This work proposes a distance that combines Minkowski and Chebyshev distances and can be seen as an intermediary distance. This combination not only achieves efficient run times in neighbourhood iteration tasks in Z 2 , but also obtains good accuracies when coupled with the k-Nearest Neighbours (k-NN) classifier. The proposed distance is approximately 1.3 times faster than Manhattan distance and 329.5 times faster than Euclidean distance in discrete neighbourhood iterations. An accuracy analysis of the k-NN classifier using a total of 33 datasets from the UCI repository, 15 distances and values assigned to k that vary from 1 to 200 is presented. In this experiment, the proposed distance obtained accuracies that were better than the average more often than its counterparts (in 26 cases out of 33), and also obtained the best accuracy more frequently (in 9 out of 33 cases). [ABSTRACT FROM AUTHOR]

Suicide by hanging is a sentinel event and a major cause of death in prisons, with an increasing frequency over recent years. The rapid detection of suicidal behavior can reduce the mortality rate and increase the odds of survival for the suicide victim. Significant efforts have been made to develop technologies for preventing hanging attempts, but most of them use cumbersome devices, or they are mainly depending on human attention and intervention. In this paper, we propose a vision-based method to automatically detect suicide by hanging. Our intelligent video surveillance system operates using depth streams provided by an RGB-D camera, regardless of illumination conditions. The proposed algorithm is based on the exploitation of the body joints’positions to model suicidal behavior. Both dynamic and static pose characteristics are calculated in order to efficiently capture the body joints’movement and model suicidal behavior. Results from the experiments on realistic video sequences, show that our system achieves a high accuracy in detecting suicide attempts, while meeting real-time requirements. [ABSTRACT FROM AUTHOR]

Video prediction plays a fundamental role in video analysis and pattern recognition. However, the generated future frames are often blurred, which are not sufficient for further research. To overcome this obstacle, this paper proposes a new deep generative video prediction network under the framework of generative adversarial nets. The network consists of three components: a motion encoder, a frame generator and a frame discriminator. The motion encoder receives multiple frame differences (also known as Eulerian motion ) as input and outputs a global video motion representation. The frame generator is a pseudo-reverse two-stream network to generate the future frame. The frame discriminator is a discriminative 3D convolution network to determine whether the given frame is derived from the true future frame distribution or not. The frame generator and frame discriminator train jointly in an adversarial manner until a Nash equilibrium. Motivated by theories on color filter array, this paper also designs a novel cross channel color gradient (3CG) loss as a guidance of deblurring. Experiments on two state-of-the-art data sets demonstrate that the proposed network is promising. [ABSTRACT FROM AUTHOR]

It is generally much harder to reconstruct the 3D structures of a non-rigid object than that of a rigid object, because of the increased degrees of freedom. To overcome the ill-posed nature of this problem, there have been many researches that incorporate prior assumptions on the object. An attractive approach is to assume the local parts to be rigid, because there might be a possibility to reduce some degrees of freedom that have to be estimated. Unfortunately, this assumption is not strictly true for general objects, and rigid and non-rigid parts are usually mixed in an object. This issue may be circumvented if we can utilize a robust local-rigidity prior to enforce rigidity on local parts selectively. In this paper, we provide a simple cost function that imposes a rigidity prior on each local triangular patch, and the proposed algorithm minimizes this cost to reconstruct a non-rigid surface. Non-rigid local parts are handled as outliers in the proposed algorithm, which makes the algorithm capable of reconstructing a near-locally-rigid surface that might contain a few non-negligible deformations. Experimental results show that the proposed method provides the state-of-the-art performance for well-known dense surface data sets. [ABSTRACT FROM AUTHOR]

Previous Spectral Feature Selection (SFS) methods output promising feature selection results in many real-world applications, which deeply depend on the preservation of the local or global structures of the data via learning a graph matrix. However, current SFS methods 1) learn the graph matrix in the original data which may contain a number of noise to affect the results of feature selection, 2) conduct the learning of a low-dimensional feature space and the graph matrix individually, thus hard achieve the optimal results of feature selection even though both of these two steps achieve their individual optimization, and 3) consider either the local or global structure of data to difficult provide complementary information for feature selection. To address the above issues, this paper proposes a novel supervised feature selection algorithm to simultaneously preserve the local structure (via adaptive structure learning in a low-dimensional feature space of the original data) and the global structure (via a low-rank constraint) of the data. Moreover, we also propose a new optimization method to fast optimize the resulting objective function. We finally verify the proposed method on eight real-word and benchmark datasets, by comparing with the state-of-the-art feature selection methods, and experimental results show that our proposed method achieves competitive results in term of classification performance. [ABSTRACT FROM AUTHOR]

So far, most of existing feature selection methods have two defects: 1) These methods are more or less heavy workload, 2) Their effect is not very good enough. To solve the above issues, a novel unsupervised feature selection algorithm is proposed. Specifically, the proposed method uses the property of the data to construct self-representation coefficient matrix, and utilizes sparse representation to find the sparse structure of the self-representation coefficient matrix, and embeds a hypergraph Laplacian regularization term to make up the insignificance of ordinary graph in the representation of multiple relations. The Linear Discriminant Analysis (LDA) algorithm is used to further adjust the result of feature selection. Finally, a low rank constraint is used to capture global structure of data. Experimental results on real datasets showed that the proposed method outperformed the state-of-the-art methods. [ABSTRACT FROM AUTHOR]

This paper studies an example-driven k -parameter computation that identifies different k values for different test samples in k NN prediction applications, such as classification, regression and missing data imputation. This is carried out with reconstructing a sparse coefficient matrix between test samples and training data. In the reconstruction process, an ℓ 1 − norm regularization is employed to generate an element-wise sparsity coefficient matrix, and an LPP (Locality Preserving Projection) regularization is adopted to keep the local structures of data for achieving the efficiency. Further, with the learnt k value, k NN approach is applied to classification, regression and missing data imputation. We experimentally evaluate the proposed approach with 20 real datasets, and show that our algorithm is much better than previous k NN algorithms in terms of data mining tasks, such as classification, regression and missing value imputation. [ABSTRACT FROM AUTHOR]

Load forecasting is an important part of power grid management. Accurate and timely load forecasting is of great significance to formulate economical and reasonable power allocation plan, improve safety and economy of power grid operation and improve power quality. In this paper, in order to find electricity load forecasting model, we propose an electricity load forecasting function mining algorithm based on artificial fish swarm and gene expression programming (ELFFM-AFSGEP). On the basis, distributed load forecast model mining based on hybrid gene expression programming and cloud computing (DLFMM-HGEPCloud) is proposed to solve the problem of massive electricity load forecasting. In order to better solve global electricity load forecasting model, error minimization crossover is introduced into DLFMM-HGEPCloud. The performance of the proposed algorithm in this paper is evaluated with a real-world dataset, and compared with GEP and some published algorithms by using the same dataset. Experimental results show that our proposed algorithm has an advantage in average time-consumption, average number of convergence, forecasted accuracy and excellent parallel performance in speedup and scaleup. [ABSTRACT FROM AUTHOR]

In this paper, we developed a novel distributed and privatized framework for miRNA-disease association prediction (DPFMDA) to predict potential miRNA-disease associations. DPFMDA assumes that diseases and miRNAs can be represented as feature matrices. Through independent factorization, collaborative analysis and matrix re-factorization, feature matrices are estimated and potential associations in miRNA-disease association matrix are predicted. The proposed computational model is tested and verified by leave-one-out cross validation (LOOCV). The AUCs of global and local LOOCV are 0.8859 and 0.8267 respectively, which outperform the state-of-the-art computational models. DPFMDA is further evaluated with case studies of four important human complex diseases. We test Esophageal Neoplasms and Prostate Neoplasms using DPFMDA in the HMDD v2.0 database. For Breast Neoplasms, we use the HMDD v2.0 database which removes all the known miRNAs-disease associations on Breast Neoplasms. For Hepatocellular Carcinoma, HMDD v1.0 database is used as the input of DPFMDA and predicted results are verified with HMDD v2.0, dbDEMC and miR2Disease databases. The results show that 90% (Esophageal Neoplasms), 88% (Prostate Neoplasms), 98% (Breast Neoplasms) and 86% (Hepatocellular Carcinoma) of top 50 predicted miRNA-disease associations are confirmed by recent experimental studies respectively. It is anticipated that DPFMDA would be an effective and extensible method for potential miRNA-disease association prediction. [ABSTRACT FROM AUTHOR]

Recently, thanks to the state-of-the-art techniques in Generative Adversarial Networks, a lot of work achieves remarkable performance on learning the mapping between an input image and an output image without any paired relation. However, traditional methods on image-to-image translation merely consider the visual appearance properties, they fail to maintain the true semantics of an image during the transfer learning procedure from source to target domain. We propose a new approach that utilizes GAN to translate unpaired images between domains and remain high level semantic abstraction aligned. Our model controls the hierarchical semantics of images by processing semantic information on label level and spatial level respectively by constructing label and attention consistent losses. The experimental results on several benchmark datasets show that generated samples are both visually similar with target images and semantically consistent with their source counterparts. Furthermore, the experiment also suggests that our method can effectively improve the classification performance in unsupervised domain adaptation problem. [ABSTRACT FROM AUTHOR]

In this paper, we propose a novel approach for ground segmentation and free space estimation of outdoor environments. The system is completely self-supervised and relies on two modules: the first module is built around a Fully Convolutional Network (FCN), and is used for ground segmentation after the system is initiated. The second module relies on depth information paired with interactive graphs cuts, and is used to train the FCN at startup, and anytime the FCN’s performance degrades during runtime. This usually happens when the camera observes a new type of outdoor scene, which is foreign to the FCN. Experiments were conducted on three datasets of different ruggedness to highlight the advantages of the proposed method. [ABSTRACT FROM AUTHOR]

Finite mixture of Gaussian regression (FMR) is a widely-used modeling technique in supervised learning problems. In cases where the number of features is large, feature selection is desirable to enhance model interpretability and to avoid overfitting. In this paper, we propose a robust feature selection method via l 2,1 -norm penalized maximum likelihood estimation (MLE) in FMR, with extension to sparse l 2,1 penalty by combining l 1 -norm with l 2,1 -norm for increasing flexibility. To solve the non-convex and non-smooth problem of (sparse) penalized MLE in FMR, we develop an new EM-based algorithm for numerical optimization, with combination of block coordinate descent and majorizing-minimization scheme in M-step. We finally apply our method in six simulations and one real dataset to demonstrate its superior performance. [ABSTRACT FROM AUTHOR]

In this work, we present a simple, robust and fast method to the perspective- n -point (P n P) problem for determining the position and orientation of a calibrated camera from known reference points. Our method transfers the pose estimation problem into an optimal problem, and only needs to solve a seventh-order and a fourth-order univariate polynomial, respectively, which makes the processes more easily understood and significantly improves the performance. Additionally, the number of solutions of the proposed method is substantially smaller than existing methods. Experiment results show that the proposed method can stably handle all 3D point configurations, including the ordinary 3D case, the quasi-singular case, and the planar case, and it offers accuracy comparable or better than that of the state-of-art methods, but at much lower computational cost. [ABSTRACT FROM AUTHOR]

Optical aberrations of a lens introduce lens blur to photographed images. Lens blur is non-stationary with the amount and characteristics of blur varying depending on spatial pixel locations in an image. This work presents non-stationary deep networks for the restoration of non-stationary lens blur. Deep networks have relatively larger receptive fields. However, the receptive fields of stationary deep networks are not wide enough for the networks to cope with the non-stationarity of lens blur that span the entire image. We use spatial pixel locations as an additional input to networks to let the network utilize location dependent features to handle the non-stationarity. Experimental results show that even shallower non-stationary networks provide better performance than deeper stationary networks. The non-stationary networks are trained from pairs of images photographed at different aperture settings, eliminating the necessity of estimation or measurement of pixel-wise variant non-stationary lens blur. [ABSTRACT FROM AUTHOR]

Estimating count and density maps from crowd images has a wide range of applications such as video surveillance, traffic monitoring, public safety and urban planning. In addition, techniques developed for crowd counting can be applied to related tasks in other fields of study such as cell microscopy, vehicle counting and environmental survey. The task of crowd counting and density map estimation is riddled with many challenges such as occlusions, non-uniform density, intra-scene and inter-scene variations in scale and perspective. Nevertheless, over the last few years, crowd count analysis has evolved from earlier methods that are often limited to small variations in crowd density and scales to the current state-of-the-art methods that have developed the ability to perform successfully on a wide range of scenarios. The success of crowd counting methods in the recent years can be largely attributed to deep learning and publications of challenging datasets. In this paper, we provide a comprehensive survey of recent Convolutional Neural Network (CNN) based approaches that have demonstrated significant improvements over earlier methods that rely largely on hand-crafted representations. First, we briefly review the pioneering methods that use hand-crafted representations and then we delve in detail into the deep learning-based approaches and recently published datasets. Furthermore, we discuss the merits and drawbacks of existing CNN-based approaches and identify promising avenues of research in this rapidly evolving field. [ABSTRACT FROM AUTHOR]

In this paper, we propose a novel semi-supervised learning framework for violence detection in video surveillance. With this framework, a classifier which distinguishes violent behavior from normal behavior can be trained using inexpensive unlabeled data with the assistance of human operators. Our approach can learn a single dictionary and a predictive linear classifier jointly. Specifically, we integrate the reconstruction error of labeled and unlabeled data, representation constraints and the coefficient incoherence into an objective function for dictionary learning, which enhances the representative and discriminative power of the established dictionary. This has contributed to that the dictionary and the classifier learned from the labeled set yield very small generalization error on unseen data. Experimental results on benchmark datasets have demonstrated the effectiveness of our approach in violence detection. [ABSTRACT FROM AUTHOR]

Learning by deep convolutional networks have shown an outstanding effectiveness in a variety of vision based classification tasks, and for which, large datasets are the prerequisites to guarantee its high performance. But in many realistic circumstances, using a massive quantity of training samples to achieve more sophisticated analysis is hard to be fulfilled always, such as human action recognition in videos, and the resulting problem of data deficiency, especially for the labeled data, would critically limit the deeper model structure as a promising solution due to its high risk of overfitting. Additionally, in lacking of high modeling capacity constrained by of model depth, the high-level visual cues like object interaction, scene context and pose variations concurrent with human action also could become the extrinsic and intrinsic challenges for the traditional deep convolutional networks. For the limitations above, in this paper, we proposed a strategy of dataset remodeling by transferring parameters of ResNet-101 layers trained on the ImageNet dataset to initialize learning model and adopt an augmented data variation approach to overcome the overfitting challenge of sample deficiency. For model structure improvement, a novel deeper two-stream ConvNets has been designed for the learning of action complexity. With a dis-order strategy of training/testing video sets, the proposed model and learning strategy are able to collaboratively achieve a significant improvement of action recognition. Experiments on two challenging datasets UCF101 and KTH have verified a superior performance in comparison with other state-of-the-art methods. [ABSTRACT FROM AUTHOR]

In this paper, we proposed a novel biometric quality assessment (BQA) method for face images and explored its applications in face recognition. Here, we considered five categories of common homogeneous distortion in video suvillance applications, i.e. low-resolution, blurring, additive Gaussian white noise, salt and pepper noise, and Poisson noise. In the BQA model, we first learnt a classifier to simultaneously predict the categories and degrees of the degradation in a face image. Because the quality labels are often ambiguous and inaccurate, we used a light convolutional neural network with the Max-Feature-Map units to make the BQA model robust to noisy labels. Afterwards, we calculated the biometric quality score by pooling such predictions based on the recognition confidence of each degradation class. Finally, we proposed one promising strategy for developing reliable face recognition systems based on this BQA method. Thorough experiments have been conducted on the CASIA, FLW, and YouTube databases. The results demonstrate the effectiveness of the proposed BQA method. [ABSTRACT FROM AUTHOR]

Cross-view gait recognition can be regarded as a domain adaption problem, in which, probe gait to be recognized in one view is different from gallery gaits collected in another view. In this paper, we present a subspace learning based method, called Multiview Max-Margin Subspace Learning (MMMSL), to learn a common subspace for associating gait data across different views. A group of projection matrices that respectively map data from different views into the common subspace are optimized via simultaneously minimizing the within-class variations and maximizing the local between-class variations of the low-dimensional embeddings from both inter-view and intra-view. In the learnt subspace, same-class samples from all views cluster together, and each different-class cluster is kept away from its nearest neighbors as far as possible. Experimental results on two benchmark gait databases, CASIA-B and OU-ISIR, demonstrate the effectiveness of the proposed method. Extensive experiments also show that our MMMSL achieves significant improvements compared with related subspace learning based methods. [ABSTRACT FROM AUTHOR]

In this paper, we propose three effective binary face descriptor learning methods, namely dual-cross patterns from three orthogonal planes (DCP-TOP), hot wheel patterns (HWP) and HWP-TOP for macro/micro-expression representation. We use feature selection to make the binary descriptors compact. Because of the limited labeled micro-expression samples, we leverage abundant labeled macro-expression and speech samples to train a more accurate classifier. Coupled metric learning algorithm is employed to model the shared features between micro-expression samples and macro-information. Smooth SVM (SSVM) is selected as a classifier to evaluate the performance of micro-expression recognition. Extensive experimental results show that our proposed methods yield the state-of-the-art classification accuracies on the CASMEII database. [ABSTRACT FROM AUTHOR]

In this paper, we propose a novel method to detect abnormal events from videos based on a maximum a posteriori (MAP). Conventional methods consider the events with low-probability with respect to a model of normal behavior as anomaly. Different from the traditional approaches, the anomaly detection is achieved by a MAP estimation in our framework. The prior knowledge is obtained from the background subtraction due to the fact that the anomalies often occur at the locations consisting of moving objects, and the likelihood function is computed by comparing the similarity between the testing samples and a designed maximum grid template. Experiments on three public databases show that our method can effectively detect abnormal events in complex scenes. [ABSTRACT FROM AUTHOR]

Crimes with respect to ATMs (Automatic Teller Machines) have attracted more and more attention, where criminals deliberately cover their faces in order to avoid being identified. This paper proposes a fast and robust face occlusion detection algorithm for ATM surveillance, which is demonstrated to be effective and efficient to handle arbitrarily occluded faces. In this algorithm, we innovatively propose to make use of the Omega shape formed by the head and shoulder of the person for head localization to tackle severe face occlusion. For this purpose, we first construct a novel energy function for elliptical head contour detection. Then, we develop a fast and robust head tracking algorithm, which utilizes the gradient and shape cues in a Bayesian framework. Lastly, to verify whether a detected head is occluded or not, we propose to fuse information from both skin color and face structure using the AdaBoost algorithm. Experimental results on real world data show that our proposed algorithm can achieve 98.64% accuracy on face detection and 98.56% accuracy on face occlusion detection, even though there are severe occlusions in faces, at a speed of up to 12 frames per second. [ABSTRACT FROM AUTHOR]

Face recognition in real-world video surveillance needs to deal with a lot of challenges including low resolution, illumination variations, pose changes, occlusions and so on. Among them, occlusions are difficult and have not attracted enough attentions. To address this problem, in this paper, we propose a robust discriminative nonnegative dictionary learning method for occluded face recognition, which estimates the occlusions adaptively and selects the features robustly. Instead of modeling the reconstruction errors using a specific distribution, we estimate occlusions adaptively according to the reconstruction errors and learn different weights for different pixels during the iterative processing. To enhance discriminant ability of the dictionary, we constrain the low-dimensional representations of samples from the same class to be as close as possible and select the discriminative features robustly via ℓ 2, 1 -norm. For the induced non-convex problem, we reformulate it into local convex optimization subproblem via utilizing the half-quadratic technique and propose new update rules. Extensive experiments are implemented on four benchmark datasets, and the experimental results demonstrate the effectiveness and robustness of the proposed method. [ABSTRACT FROM AUTHOR]

Face recognition research has benefited from the availability of challenging face databases and benchmark results on popular databases show very high performance on single-person per image/video databases. However, in real world surveillance scenarios, the environment is unconstrained and the videos are likely to record multiple subjects within the field of view. In such crowd surveillance videos, both face detection and recognition are still considered as onerous tasks. One of the key factors for limited research in this direction is unavailability of benchmark databases. This paper presents CrowdFaceDB video face database that fills the gap in unconstrained face recognition for crowd surveillance. The two fold contributions are: (1) developing an unconstrained crowd video face database of over 250 subjects, and (2) creating a benchmark protocol and performing baseline experiments for both face detection and verification. The experimental results showcase the exigent nature of crowd surveillance and limitations of existing algorithms/systems. [ABSTRACT FROM AUTHOR]

Distance metric learning has been widely applied for person re-identification. However, the typical Small Sample Size (SSS) problem, which is induced by high dimensional feature and limited training samples in most re-identification datasets, may lead to a sub-optimal metric. In this work, we propose to embed samples into a discriminative null space based on Marginal Fisher Analysis (MFA) to overcome the SSS problem. In such a null space, multiple images of the same pedestrian are collapsed to a single point, resulting the extreme Marginal Fisher Criterion. We theoretically analyze the null space and derive its closed-form solution which can be computed very efficiently. To deal with the heavy storage burden in computation, we further extend the proposed method to kernel version, which is called Kernel Null Space Marginal Fisher Analysis (KNSMFA). Experiments on four challenging person re-identification datasets show that KNSMFA uniformly outperforms state-of-the-art approaches. [ABSTRACT FROM AUTHOR]

In recent years, there has been increasing demand for social security and identity authentication, which leads to the booming of many biometrics involved large-scale problems such as recognition, retrieval, and identification. In this case, traditional models are infeasible due to the limited capability for handling large-scale data. Therefore, hashing technique is becoming prevalent due to its low storage cost and fast query speed. Recently, researchers have shown that supervised hashing can achieve higher accuracy than unsupervised hashing by incorporating tag or label information of data for learning hashing function. However, existing supervised methods treat all training examples equally, ignoring the different impacts of various training examples on the learning process. As a result, their performance is not satisfactory under some practical situations. As an improvement, this paper proposes a new method called “Supervised Hashing with Informative Set Selection” (SHISS), which assumes that different training examples have different influence on the learning process, and their usage should follow a logic way during optimization. In particular, we propose two criteria, certainty and diversity, to evaluate the informativeness of the subsets of training examples and encourage the more informative subsets to be learned ahead of the less informative ones. The experiments on two typical image retrieval datasets and one face dataset demonstrate that the proposed SHISS obtains higher mean average precision and shows faster convergence rate compared with the state-of-the-art hashing methods. [ABSTRACT FROM AUTHOR]

Rough set theory has been playing a significant role in data mining, but deficiencies still exist in fast attribute reduction and self-learning for decision systems. Grid-clustered rough set (GCRS) model computing the universe partition based on grid subspace cluster (GSC) algorithm is constructed in this paper to fill the aforementioned gaps. GSC algorithm characterized by densities and distances in grid subspace is firstly presented along with the automatically selecting of cluster centers. Then a rough self-learning theory including extensional learning and intensional learning is raised. To realize quick attribute reduction and self-learning, a novel rough set model named GCRS that integrates rough set theory and GSC algorithm is proposed, and a corresponding quick attribute reduction algorithm and a rough self-learning algorithm based on GCRS are subsequently designed. A multitude of experiments demonstrate three aspects: a) results of attribute reduction using GCRS model are proved to be of higher classification accuracies; b) attribute reduction based on GCRS model for big data manifests faster speed; c) results of rough self-learning experiments illustrate the validity of the proposed theory. [ABSTRACT FROM AUTHOR]

Deep convolutional neural networks have made great progress in recent handwritten character recognition (HCR) by learning discriminative features from large amounts of labeled data. However, the large variance of handwriting styles across writers is still a big challenge to the robust HCR. To alleviate this issue, an intuitional idea is to extract writer-independent semantic features from handwritten characters, while standard printed characters are writer-independent stencils for handwritten characters. They could be used as prior knowledge to guide models to exploit writer-independent semantic features for HCR. In this paper, we propose a novel adversarial feature learning (AFL) model to incorporate the prior knowledge of printed data and writer-independent semantic features to improve the performance of HCR on limited training data. Different from available handcrafted features methods, the proposed AFL model exploits writer-independent semantic features automatically, and standard printed data as prior knowledge is learnt objectively. Systematic experiments on MNIST and CASIA–HWDB show that the proposed model is competitive with the state-of-the-art methods on the offline HCR task. [ABSTRACT FROM AUTHOR]

Keypoint detection usually results in a large number of keypoints which are mostly clustered, redundant, and noisy. These keypoints often require special processing like Adaptive Non-Maximal Suppression (ANMS) to retain the most relevant ones. In this paper, we present three new efficient ANMS approaches which ensure a fast and homogeneous repartition of the keypoints in the image. For this purpose, a square approximation of the search range to suppress irrelevant points is proposed to reduce the computational complexity of the ANMS. To further speed up the proposed approaches, we also introduce a novel strategy to initialize the search range based on image dimension which leads to a faster convergence. An exhaustive survey and comparisons with already existing methods are provided to highlight the effectiveness and scalability of our methods and the initialization strategy. [ABSTRACT FROM AUTHOR]

This paper is concerned with the fundamentals of integrating crowd knowledge such as ratings, opinions or tags provided by the internet users. As a concrete example, we consider the problem of image recognition based on user-provided tags. Each user is assumed to have certain knowledge about the images, which can be incomplete or only of partial relevance to the recognition task. Each user is also assumed to have his own choice of tag vocabulary, possibly different from the set of prescribed labels for image recognition. We argue that a user’s knowledge can be separated into the structure of the knowledge and the representation of the structure (namely, his tag vocabulary). This perspective advocates a systematic three-step methodology for crowd knowledge integration in such applications, whereby the problem of interest is decoupled into three sub-problems in tandem: knowledge structure aggregation, vocabulary interpretation, and label assignment. We derive a lower bound for the achievable error probability. Using this bound and via Monte-Carlo simulations, we investigate the performance of a knowledge integration system in relation to various parameter settings. [ABSTRACT FROM AUTHOR]

This paper addresses the problem of unsupervised object localization in an image. Unlike previous supervised and weakly supervised algorithms that require bounding box or image level annotations for training classifiers, we propose a simple yet effective technique for localization using iterative spectral clustering. This iterative spectral clustering approach along with appropriate cluster selection strategy in each iteration naturally helps in searching of object region in the image. In order to estimate the final localization window, we group the proposals obtained from the iterative spectral clustering step based on the perceptual similarity, and average the coordinates of the proposals from the top scoring groups. We benchmark our algorithm on challenging datasets like Object Discovery and PASCAL VOC 2007, achieving an average CorLoc percentage of 51% and 35% respectively which is comparable to various other weakly supervised algorithms despite being completely unsupervised. [ABSTRACT FROM AUTHOR]

Keeping record of daily meal intake is an effective solution for tackling with obesity and overweight. This can be done by developing apps on smartphones that are able to automatically recommend a short list of most probable foods by analyzing the photo taken from food. Then, the user chooses the correct answer from the short list. Hence, the automatic food recognition system must be able to recommend an accurate list. In other words, it is not essential for these apps to have a very high top-1 accuracy. Considering that the app will show the list of 5 most probable foods, the food recognition system must have a high top-5 accuracy. A food recognition system is usually developed by adapting knowledge of state-of-the-art networks such as GoogleNet and ResNet to the domain of food. However, these networks have high number of parameters. In this paper, we propose a 23-layer architecture which has 99.14% and 96.63% fewer parameter compared with ResNet and GoogleNet. Our experiment on Food101 and UECFood-256 datasets shows that although our network reduces the number of parameters dramatically, it produces more accurate results than GoogleNet and its accuracy is comparable with ResNet. [ABSTRACT FROM AUTHOR]