Feature Space Representation Research Articles

Investigation of Event-Based Surfaces for High-Speed Detection, Unsupervised Feature Extraction, and Object Recognition.

In this work, we investigate event-based feature extraction through a rigorous framework of testing. We test a hardware efficient variant of Spike Timing Dependent Plasticity (STDP) on a range of spatio-temporal kernels with different surface decaying methods, decay functions, receptive field sizes, feature numbers, and back end classifiers. This detailed investigation can provide helpful insights and rules of thumb for performance vs. complexity trade-offs in more generalized networks, especially in the context of hardware implementation, where design choices can incur significant resource costs. The investigation is performed using a new dataset consisting of model airplanes being dropped free-hand close to the sensor. The target objects exhibit a wide range of relative orientations and velocities. This range of target velocities, analyzed in multiple configurations, allows a rigorous comparison of time-based decaying surfaces (time surfaces) vs. event index-based decaying surface (index surfaces), which are used to perform unsupervised feature extraction, followed by target detection and recognition. We examine each processing stage by comparison to the use of raw events, as well as a range of alternative layer structures, and the use of random features. By comparing results from a linear classifier and an ELM classifier, we evaluate how each element of the system affects accuracy. To generate time and index surfaces, the most commonly used kernels, namely event binning kernels, linearly, and exponentially decaying kernels, are investigated. Index surfaces were found to outperform time surfaces in recognition when invariance to target velocity was made a requirement. In the investigation of network structure, larger networks of neurons with large receptive field sizes were found to perform best. We find that a small number of event-based feature extractors can project the complex spatio-temporal event patterns of the dataset to an almost linearly separable representation in feature space, with best performing linear classifier achieving 98.75% recognition accuracy, using only 25 feature extracting neurons.

Multi-Pose Facial Expression Recognition Based on Generative Adversarial Network

The recognition of human emotions from facial expression images is one of the most important topics in the machine vision and image processing fields. However, recognition becomes difficult when dealing with non-frontal faces. To alleviate the influence of poses, we propose an encoder-decoder generative adversarial network that can learn pose-invariant and expression-discriminative representations. Specifically, we assume that a facial image can be divided into an expressive component, an identity component, a head pose component and a remaining component. The encoder encodes each component into a feature representation space and the decoder recovers the original image from these encoded features. A classification loss on the components and an l 1 pixel-wise loss are applied to guarantee the rebuilt image quality and produce more constrained visual representations. Quantitative and qualitative evaluations on two multi-pose datasets demonstrate that the proposed algorithm performs favorably compared to state-of-the-art methods.

Tonic-clonic seizure detection using accelerometry-based wearable sensors: A prospective, video-EEG controlled study

PurposeThe aim of this prospective, video-electroencephalography (video-EEG) controlled study was to evaluate the performance of an accelerometry-based wearable system to detect tonic-clonic seizures (TCSs) and to investigate the accuracy of different seizure detection algorithms using separate training and test data sets. MethodsSeventy-five epilepsy surgery candidates undergoing video-EEG monitoring were included. The patients wore one three-axis accelerometer on each wrist during video-EEG. The accelerometer data was band-pass filtered and reduced using a movement threshold and mapped to a time-frequency feature space representation. Algorithms based on standard binary classifiers combined with a TCS specific event detection layer were developed and trained using the training set. Their performance was evaluated in terms of sensitivity and false positive (FP) rate using the test set. ResultsThirty-seven available TCSs in 11 patients were recorded and the data was divided into disjoint training (27 TCSs, three patients) and test (10 TCSs, eight patients) data sets. The classification algorithms evaluated were K-nearest-neighbors (KNN), random forest (RF) and a linear kernel support vector machine (SVM). For the TCSs detection performance of the three algorithms in the test set, the highest sensitivity was obtained for KNN (100% sensitivity, 0.05 FP/h) and the lowest FP rate was obtained for RF (90% sensitivity, 0.01 FP/h). ConclusionsThe low FP rate enhances the clinical utility of the detection system for long-term reliable seizure monitoring. It also allows a possible implementation of an automated TCS detection in free-living environment, which could contribute to ascertain seizure frequency and thereby better seizure management.

A comprehensive overview of feature representation for biometric recognition

The performance of any biometric recognition system heavily dependents on finding a good and suitable feature representation space where observations from different classes are well separated. Unfortunately, finding this proper representation is a challenging problem which has taken a huge interest in machine learning and computer vision communities. In the this paper we present a comprehensive overview of the different existing feature representation techniques. This is carried out by introducing simple and clear taxonomies as well as effective explanation of the prominent techniques. This is intended to guide the neophyte and provide researchers with state-of-the-art approaches in order to help advance the research topic in biometrics.

Using Deep Convolutional Neural Network Architectures for Object Classification and Detection Within X-Ray Baggage Security Imagery

We consider the use of deep convolutional neural networks (CNNs) with transfer learning for the image classification and detection problems posed within the context of X-ray baggage security imagery. The use of the CNN approach requires large amounts of data to facilitate a complex end-to-end feature extraction and classification process. Within the context of X-ray security screening, limited availability of object of interest data examples can thus pose a problem. To overcome this issue, we employ a transfer learning paradigm such that a pre-trained CNN, primarily trained for generalized image classification tasks where sufficient training data exists, can be optimized explicitly as a later secondary process towards this application domain. To provide a consistent feature-space comparison between this approach and traditional feature space representations, we also train support vector machine (SVM) classifier on CNN features. We empirically show that fine-tuned CNN features yield superior performance to conventional hand-crafted features on object classification tasks within this context. Overall we achieve 0.994 accuracy based on AlexNet features trained with SVM classifier. In addition to classification, we also explore the applicability of multiple CNN driven detection paradigms, such as sliding window-based CNN (SW-CNN), Faster region-based CNNs (F-RCNNs), region-based fully convolutional networks (R-FCN), and YOLOv2. We train numerous networks tackling both single and multiple detections over SW-CNN/ F-RCNN/R-FCN/YOLOv2 variants. YOLOv2, Faster-RCNN, and R-FCN provide superior results to the more traditional SW-CNN approaches. With the use of YOLOv2, using input images of size $544\times 544$ , we achieve 0.885 mean average precision (mAP) for a six-class object detection problem. The same approach with an input of size $416\times 416$ yields 0.974 mAP for the two-class firearm detection problem and requires approximately 100 ms per image. Overall we illustrate the comparative performance of these techniques and show that object localization strategies cope well with cluttered X-ray security imagery, where classification techniques fail.

Deep UQ: Learning deep neural network surrogate models for high dimensional uncertainty quantification

State-of-the-art computer codes for simulating real physical systems are often characterized by vast number of input parameters. Performing uncertainty quantification (UQ) tasks with Monte Carlo (MC) methods is almost always infeasible because of the need to perform hundreds of thousands or even millions of forward model evaluations in order to obtain convergent statistics. One, thus, tries to construct a cheap-to-evaluate surrogate model to replace the forward model solver. For systems with large numbers of input parameters, one has to address the curse of dimensionality through suitable dimensionality reduction techniques. A popular class of dimensionality reduction methods are those that attempt to recover a low-dimensional representation of the high-dimensional feature space. However, such methods often tend to overestimate the intrinsic dimensionality of the input feature space. In this work, we demonstrate the use of deep neural networks (DNN) to construct surrogate models for numerical simulators. We parameterize the structure of the DNN in a manner that lends the DNN surrogate the interpretation of recovering a low-dimensional nonlinear manifold. The model response is a parameterized nonlinear function of the low-dimensional projections of the input. We think of this low-dimensional manifold as a nonlinear generalization of the notion of the active subspace. Our approach is demonstrated with a problem on uncertainty propagation in a stochastic elliptic partial differential equation (SPDE) with uncertain diffusion coefficient. We deviate from traditional formulations of the SPDE problem by lifting the assumption of fixed lengthscales of the uncertain diffusion field. Instead we attempt to solve a more challenging problem of learning a map between an arbitrary snapshot of the diffusion field and the response.

Feature Extraction by Using Dual-Generalized Discriminative Common Vectors

In this paper, a dual online subspace-based learning method called dual-generalized discriminative common vectors (Dual-GDCV) is presented. The method extends incremental GDCV by exploiting simultaneously both the concepts of incremental and decremental learning for supervised feature extraction and classification. Our methodology is able to update the feature representation space without recalculating the full projection or accessing the previously processed training data. It allows both adding information and removing unnecessary data from a knowledge base in an efficient way, while retaining the previously acquired knowledge. The proposed method has been theoretically proved and empirically validated in six standard face recognition and classification datasets, under two scenarios: (1) removing and adding samples of existent classes, and (2) removing and adding new classes to a classification problem. Results show a considerable computational gain without compromising the accuracy of the model in comparison with both batch methodologies and other state-of-art adaptive methods.

Model-guided search for optimal natural-science-category training exemplars: A work in progress.

Under the guidance of a formal exemplar model of categorization, we conduct comparisons of natural-science classification learning across four conditions in which the nature of the training examples is manipulated. The specific domain of inquiry is rock classification in the geologic sciences; the goal is to use the model to search for optimal training examples for teaching the rock categories. On the positive side, the model makes a number of successful predictions: Most notably, compared with conditions involving focused training on small sets of training examples, generalization to novel transfer items is significantly enhanced in a condition in which learners experience a broad swath of training examples from each category. Nevertheless, systematic departures from the model predictions are also observed. Further analyses lead us to the hypothesis that the high-dimensional feature-space representation derived for the rock stimuli (to which the exemplar model makes reference) systematically underestimates within-category similarities. We suggest that this limitation is likely to arise in numerous situations in which investigators attempt to build detailed feature-space representations for naturalistic categories. A low-parameter extended version of the model that adjusts for this limitation provides dramatically improved accounts of performance across the four conditions. We outline future steps for enhancing the current feature-space representation and continuing our goal of using formal psychological models to guide the search for effective methods of teaching science categories.

Sparse feature space representation: A unified framework for semi-supervised and domain adaptation learning

In a semi-supervised domain adaptation (DA) task, one has access to only few labeled target examples. In this case, the success of DA needs the effective utilization of a large number of unlabeled target data to extract more discriminative information that is useful for generalization. To this end, we exploit in this paper the feature space embeddings of the target data as well as multi-source prior models to augment the discrimination space for the target function learning. Therefore, we propose a novel multi-source adaptation learning framework based on Sparse Feature Space Representation (SFSR), or called SFSR-MSAL for short. Specifically, the SFSR algorithm is first presented for the further construction of robust graph, on which the discriminative information can be smoothly propagated into the unlabeled target data by additionally incorporating the geometric structure of the target data. Considering the robustness in the semi-supervised DA, we replace the traditional l2-norm based least squares regression with the l2, 1-norm sparse regression, and then construct the SFSR-graph based semi-supervised DA framework with multi-source adaptation constraints. Our framework is universal and can be easily degraded into semi-supervised learning by just tuning the regularization parameter. Moreover, to select the discriminative SFSR-graph Laplacians, we also introduce the ensemble SFSR-graph Laplacians regularization into SFSR-MSAL, thus further improving the performance of SFSR-MSAL. The validity of our methods including semi-supervised and DA learning are examined by several visual recognition tasks on some benchmark datasets, which demonstrate the superiority of our methods in comparison with other related state-of-the-art algorithms.

A Formal Psychological Model of Classification Applied to Natural-Science Category Learning

The field of psychological science has seen major advances in the development of formal models of perceptual classification learning; however, little work has tested such models in real-world natural-category domains. In our current project, we aim to fill that gap by testing the ability of a formal exemplar model of classification to predict learning of rock categories in the geologic sciences. As a prerequisite for testing the model in this domain, we have conducted extensive work to derive a high-dimensional feature-space representation for the rock stimuli. An eight-dimensional representation yields good accounts of naive participants’ judgments of similarity among a large battery of rock-picture samples; furthermore, the eight dimensions have natural psychological interpretations. We then use the exemplar model in combination with the derived feature-space representation to successfully predict participants’ learning and generalization of a variety of scientifically defined rock categories. We discuss further steps for making use of the model and its associated feature-space representation to search for effective techniques of teaching categories in the science classroom.

Tests of an exemplar-memory model of classification learning in a high-dimensional natural-science category domain.

Experiments were conducted in which novice participants learned to classify pictures of rocks into real-world, scientifically defined categories. The experiments manipulated the distribution of training instances during an initial study phase, and then tested for correct classification and generalization performance during a transfer phase. The similarity structure of the to-be-learned categories was also manipulated across the experiments. A low-parameter version of an exemplar-memory model, used in combination with a high-dimensional feature-space representation for the rock stimuli, provided good overall accounts of the categorization data. The successful accounts included (a) predicting how performance on individual item types within the categories varied with the distributions of training examples, (b) predicting the overall levels of classification accuracy across the different rock categories, and (c) predicting the patterns of between-category confusions that arose when classification errors were made. The work represents a promising initial step in scaling up the application of formal models of perceptual classification learning to complex natural-category domains. We discuss further steps for making use of the model and its associated feature-space representation to search for effective techniques of teaching categories in the science classroom. (PsycINFO Database Record

A clutter-calibrated Hough transform for the estimation of directional structure and dominant directions in grey-level images

Increasing amounts of image data are being routinely collected as part of the big-data revolution, with applications as diverse as automated security surveillance and dynamic medical imaging. To make best use of the data, the analyses must be automatic and rapid. Simple image properties can be used to highlight specific features in an initial screening or form input to elaborate classification techniques. A key stage in any image analysis is the identification of structure amongst the noise. It is important to realise that noise can be localized, independent and random, or it could contain small-scale structure which, in some ways, resembles the important features---this is called clutter. This paper uses the concept of the Hough transform to convert grey-level images into a more useful feature space representation. This space is searched for high density regions to identify dominant structure whilst taking into account micro-line clutter. Further, a directional distribution is introduced and a resulting dominant direct is proposed as a single structural summary. Many examples of simulated and real data images are used to illustrate the proposed techniques.

Weighted Feature Space Representation with Kernel for Image Classification

The kernel method is a very effective and popular method to extract features from data such as images. A novel method is presented to enhance traditional kernel method for face image representation in this paper, which is very suitable to treat the high-dimensional datasets. The proposed method is called weighted kernel representation-based method (WKRBM) in this paper. WKRBM assumes that the test sample can be expressed by all the training samples and linear solution in the mapping space. It uses the obtained linear combination to recognize face images. In particular, the coefficients of a linear combination can be set as the optimal weight that is an important factor to obtain better performance for image classification. The rationale, characteristics, and advantages of the proposed method are presented. The analysis describes that WKRBM outperforms collaborative representation-based kernel method for image recognition. Extensive experimental results illustrate that WKRBM has partial properties of sparsity, which is effective to recognize images.

Range profile target recognition using sparse representation based on feature space

A novel method is presented to improve the recognition rate of warhead in this paper. Firstly, a tool for electromagnetic calculation, like CST Microwave Studio, is used to simulate the frequency response of the electromagnetic scattering. Secondly, the echo and further the range profile are acquired from the frequency response by further processing. Thirdly, a set of discriminative features is extracted from the range profiles of the target. Fourthly, these features are used to construct a dictionary for the sparse representation classifier. Finally, the sample of the target can be classified by solving the sparsest coefficients. Since the reconstruction result is determined by a linear combination of the training samples, this method has a good robustness for the variable features. By formulating the problem within a feature-based sparse representation framework, the presented method combines the discriminative features of each sample during the sparse recovery process rather than in a postprocessing manner. Moreover, based on the feature representation space rather than a single feature or image pixel, the constructed dictionary exhibits both strong expressive and discriminative powers that can enhance the classification performance of the test sample. A series of test results based on the simulated data demonstrates the effectiveness of our method.

Toward the development of a feature-space representation for a complex natural category domain.

This article reports data sets aimed at the development of a detailed feature-space representation for a complex natural category domain, namely 30 common subtypes of the categories of igneous, metamorphic, and sedimentary rocks. We conducted web searches to develop a library of 12 tokens each of the 30 subtypes, for a total of 360 rock pictures. In one study, subjects provided ratings along a set of 18 hypothesized primary dimensions involving visual characteristics of the rocks. In other studies, subjects provided similarity judgments among pairs of the rock tokens. Analyses are reported to validate the regularity and information value of the dimension ratings. In addition, analyses are reported that derive psychological scaling solutions from the similarity-ratings data and that interrelate the derived dimensions of the scaling solutions with the directly rated dimensions of the rocks. The stimulus set and various forms of ratings data, as well as the psychological scaling solutions, are made available on an online website (https://osf.io/w64fv/) associated with the article. The study provides a fundamental data set that should be of value for a wide variety of research purposes, including: (1) probing the statistical and psychological structure of a complex natural category domain, (2) testing models of similarity judgment, and (3) developing a feature-space representation that can be used in combination with formal models of category learning to predict classification performance in this complex natural category domain.

SDRNF: generating scalable and discriminative random nonlinear features from data

Real world data analysis problems often require nonlinear methods to get successful prediction. Kernel methods, e.g. Kernelized Principal Component Analysis, are a common way to get nonlinear properties based on linear representations in a high-dimensional feature space. Unfortunately, traditional kernel methods are unscalable for large-size or even medium-size data. On the other hand, randomized algorithms have been recently proposed to extract nonlinear features in kernel methods. Compared with exact kernel methods, this family of approaches is capable of speeding up the training process dramatically, while maintaining acceptable the classification accuracy. However, these methods fail to engage discriminative features. This significantly limits their classification accuracy. In this paper, we propose a scalable and approximate technique called SDRNF for introducing both nonlinear and discriminative features based on randomized methods. By combining randomized kernel approximation with a couple of generalized eigenvector problems, the proposed approach proves both scalable and accurate for large-scale data. A series of experiments on two benchmark data sets MNIST and CIFAR-10 reveal that our method is fast and scalable, and also generates better classification accuracy over other competitive kernel approximation methods.

An empirical study on budget-aware online kernel algorithms for streams of graphs

Kernel methods are considered as an effective technique for on-line learning. Many approaches have been developed for compactly representing the dual solution of a kernel method when the problem imposes memory constraints. However, in the literature no work is specifically tailored to streams of graphs. Motivated by the fact that the size of the feature space representation of many state-of-the-art graph kernels is relatively small and thus it is explicitly computable, we study whether executing kernel algorithms in the feature space can be more effective than the classical dual approach. We study three different algorithms and various strategies for managing the budget. Efficiency and efficacy of the proposed approaches are experimentally assessed on relatively large graph streams exhibiting concept drift. It turns out that, when strict memory budget constraints have to be enforced, working in feature space, given the current state-of-the-art on graph kernels, is more than a viable alternative to dual approaches, both in terms of speed and classification performance.

Variable Star Signature Classification using Slotted Symbolic Markov Modeling

With the advent of digital astronomy, new benefits and new challenges have been presented to the modern day astronomer. No longer can the astronomer rely on manual processing, instead the profession as a whole has begun to adopt more advanced computational means. This paper focuses on the construction and application of a novel time-domain signature extraction methodology and the development of a supporting supervised pattern classification algorithm for the identification of variable stars. A methodology for the reduction of stellar variable observations (time-domain data) into a novel feature space representation is introduced. The methodology presented will be referred to as Slotted Symbolic Markov Modeling (SSMM) and has a number of advantages which will be demonstrated to be beneficial; specifically to the supervised classification of stellar variables. It will be shown that the methodology outperformed a baseline standard methodology on a standardized set of stellar light curve data. The performance on a set of data derived from the LINEAR dataset will also be shown.

Clustering of cell populations in flow cytometry data using a combination of Gaussian mixtures

We propose a supervised learning approach to automatic quantification of cell populations in flow cytometric samples. One sample contains up to millions of measurement vectors with a dimensionality between 10 and 20. Normally, each measurement vector corresponds to a single cell in the biological sample. Identifying biologically meaningful cell populations is essentially a clustering problem, however, standard clustering methods are impractical, because size, shape and location of corresponding clusters may vary strongly between samples mainly due to phenotypic differences and inter-laboratory variations. In our holistic approach, we implicitly employ the structural information (such as relative locations and shape of sub-populations). A new input sample is reconstructed by a linear combination of artificial reference samples each represented by a Gaussian Mixture Model (GMM), in which for each Gaussian component the class label of the corresponding cluster of observations is known. The reference samples are calculated from a larger set of training samples by non-negative matrix factorization and can be regarded as the basis of a lower dimensional feature space, in which input samples are reconstructed. We show a method for calculating the feature space transformation based on minimization the L2 distance defined between two GMM. The feature space representation of the sample is then used to assign each observation to one of the specified sub-populations by a Bayes decision. We present classification results on a database of about 170 patients with Acute Lymphoblastic Leukemia (ALL), where high accuracy in the prediction of relatively small leukemic populations is crucial. The approach is not limited to our application. It can be employed wherever analysis of large, multi-dimensional, numerical data of a specific class of samples with related structure has to be performed.

Accurate and efficient shape matching approach using vocabularies of multi-feature space representations

Selection of compressed, robust and accurate features is the fundamental ingredient of effective content-based image recognition and retrieval using shape information of objects in the image. In this paper, we present a four-stage system for real-time object recognition and retrieval that employs multiple feature space representation using contour information. In the first stage, we pre-process the shapes to cater for the presence of distortions such as cracks that can significantly distort the contour information of the shape. We then generate multiple feature space representations of shapes to be used later in proposed combination for efficient and accurate retrieval of shapes using a hierarchical indexing structure. To enable real-time image-based shape analysis by enhancing the efficiency and reducing the storage requirement of proposed shape descriptors, we present a quantization approach to generate vocabulary of feature space representation of shapes. These features are then combined in an ensemble for accurate and efficient shape retrieval and recognition in the presence of large shape datasets. The proposed system is evaluated using publicly available shape datasets such as MPEG 7, Swedish leaf and KIMIA 99 datasets. Our approach achieves higher accuracies which are better than state-of-the-art approaches reported in literature whilst looking at a small subset of shapes in dataset.