Mutual Information Regularized Feature-Level Frankenstein for Discriminative Recognition.

Deep learning recognition approaches can potentially perform better if we can extract a discriminative representation that controllably separates nuisance factors. In this paper, we propose a novel approach to explicitly enforce the extracted discriminative representation d, extracted latent variation l (e,g., background, unlabeled nuisance attributes), and semantic variation label vector s (e.g., labeled expressions/pose) to be independent and complementary to each other. We can cast this problem as an adversarial game in the latent space of an auto-encoder. Specifically, with the to-be-disentangled s, we propose to equip an end-to-end conditional adversarial network with the ability to decompose an input sample into d and l. However, we argue that maximizing the cross-entropy loss of semantic variation prediction from d is not sufficient to remove the impact of s from d, and that the uniform-target and entropy regularization are necessary. A collaborative mutual information regularization framework is further proposed to avoid unstable adversarial training. It is able to minimize the differentiable mutual information between the variables to enforce independence. The proposed discriminative representation inherits the desired tolerance property guided by prior knowledge of the task. Our proposed framework achieves top performance on diverse recognition tasks, including digits classification, large-scale face recognition on LFW and IJB-A datasets, and face recognition tolerant to changes in lighting, makeup, disguise, etc.

Correlation filters with controlled scale response.

Correlation filtering methods are becoming increasingly popular for image recognition and location. The recent introduction of optimal tradeoff circular harmonic function filters allowed the user to specify the response of a correlation filter to in-plane rotation distortion. In this paper we introduce a new correlation filter design that can provide a user-specified response to in-plane scale distortion. The design is based on the Mellin radial harmonic (MRH) transform and incorporates multiple harmonics into the correlation filter for improved discrimination capability. Additionally, the filter design minimizes the average correlation energy in order to achieve sharp correlation peaks, and thus we refer to these filters as minimum average correlation energy Mellin radial harmonic (MACE-MRH) filters. We present underlying theory, a MACE-MRH filter design method, and numerical simulation results.

An Automatic Subject-Adaptable Heartbeat Classifier Based on Multiview Learning.

In this paper, a novel subject-adaptable heartbeat classification model is presented, in order to address the significant interperson variations in ECG signals. A multiview learning approach is proposed to automate subject adaptation using a small amount of unlabeled personal data, without requiring manual labeling. The designed subject-customized models consist of two models, namely, general classification model and specific classification model. The general model is trained using similar subjects out of a population dataset, where a pattern matching based algorithm is developed to select the subjects that are "similar" to the particular test subject for model training. In contrast, the specific model is trained mainly on a small amount of high-confidence personal dataset, resulting from multiview-based learning. The learned general model represents the population knowledge, providing an interperson perspective for classification, while the specific model corresponds to the specific knowledge of the subject, offering an intraperson perspective for classification. The two models supplement each other and are combined to achieve improved personalized ECG analysis. The proposed methods have been validated on the MIT-BIH Arrhythmia Database, yielding an average classification accuracy of 99.4% for ventricular ectopic beat class and 98.3% for supraventricular ectopic beat class, which corresponds to a significant improvement over other published results.

Zero-Aliasing Correlation Filters for Object Recognition.

Correlation filters (CFs) are a class of classifiers that are attractive for object localization and tracking applications. Traditionally, CFs have been designed in the frequency domain using the discrete Fourier transform (DFT), where correlation is efficiently implemented. However, existing CF designs do not account for the fact that the multiplication of two DFTs in the frequency domain corresponds to a circular correlation in the time/spatial domain. Because this was previously unaccounted for, prior CF designs are not truly optimal, as their optimization criteria do not accurately quantify their optimization intention. In this paper, we introduce new zero-aliasing constraints that completely eliminate this aliasing problem by ensuring that the optimization criterion for a given CF corresponds to a linear correlation rather than a circular correlation. This means that previous CF designs can be significantly improved by this reformulation. We demonstrate the benefits of this new CF design approach with several important CFs. We present experimental results on diverse data sets and present solutions to the computational challenges associated with computing these CFs. Code for the CFs described in this paper and their respective zero-aliasing versions is available at http://vishnu.boddeti.net/projects/correlation-filters.html.

Detecting diabetes mellitus and nonproliferative diabetic retinopathy using tongue color, texture, and geometry features.

Diabetes mellitus (DM) and its complications leading to diabetic retinopathy (DR) are soon to become one of the 21st century's major health problems. This represents a huge financial burden to healthcare officials and governments. To combat this approaching epidemic, this paper proposes a noninvasive method to detect DM and nonproliferative diabetic retinopathy (NPDR), the initial stage of DR based on three groups of features extracted from tongue images. They include color, texture, and geometry. A noninvasive capture device with image correction first captures the tongue images. A tongue color gamut is established with 12 colors representing the tongue color features. The texture values of eight blocks strategically located on the tongue surface, with the additional mean of all eight blocks are used to characterize the nine tongue texture features. Finally, 13 features extracted from tongue images based on measurements, distances, areas, and their ratios represent the geometry features. Applying a combination of the 34 features, the proposed method can separate Healthy/DM tongues as well as NPDR/DM-sans NPDR (DM samples without NPDR) tongues using features from each of the three groups with average accuracies of 80.52% and 80.33%, respectively. This is on a database consisting of 130 Healthy and 296 DM samples, where 29 of those in DM are NPDR.

A framework for binding and retrieving class-specific information to and from image patterns using correlation filters.

We describe a template-based framework to bind class-specific information to a set of image patterns and retrieve that information by matching the template to a query pattern of the same class. This is done by mapping the class-specific information to a set of spatial translations which are applied to the set of image patterns from which a template is designed, taking advantage of the properties of correlation filters. The bound information is retrieved during matching with an authentic query by estimating the spatial translations applied to the images that were used to design the template. In this paper, we focus on the problem of binding information to biometric signatures as an application of our framework. Our framework is flexible enough to allow spreading the information to be bound over multiple pattern classes which, in the context of biometric key-binding, enables multiclass and multimodal biometric key-binding. We demonstrate the effectiveness of the proposed scheme via extensive numerical results on multiple biometric databases.

Maximum Margin Correlation Filter: a new approach for localization and classification.

Support vector machine (SVM) classifiers are popular in many computer vision tasks. In most of them, the SVM classifier assumes that the object to be classified is centered in the query image, which might not always be valid, e.g., when locating and classifying a particular class of vehicles in a large scene. In this paper, we introduce a new classifier called Maximum Margin Correlation Filter (MMCF), which, while exhibiting the good generalization capabilities of SVM classifiers, is also capable of localizing objects of interest, thereby avoiding the need for image centering as is usually required in SVM classifiers. In other words, MMCF can simultaneously localize and classify objects of interest. We test the efficacy of the proposed classifier on three different tasks: vehicle recognition, eye localization, and face classification. We demonstrate that MMCF outperforms SVM classifiers as well as well known correlation filters.

Heartbeat classification using morphological and dynamic features of ECG signals.

In this paper, we propose a new approach for heartbeat classification based on a combination of morphological and dynamic features. Wavelet transform and independent component analysis (ICA) are applied separately to each heartbeat to extract morphological features. In addition, RR interval information is computed to provide dynamic features. These two different types of features are concatenated and a support vector machine classifier is utilized for the classification of heartbeats into one of 16 classes. The procedure is independently applied to the data from two ECG leads and the two decisions are fused for the final classification decision. The proposed method is validated on the baseline MIT-BIH arrhythmia database and it yields an overall accuracy (i.e., the percentage of heartbeats correctly classified) of 99.3% (99.7% with 2.4% rejection) in the "class-oriented" evaluation and an accuracy of 86.4% in the "subject-oriented" evaluation, comparable to the state-of-the-art results for automatic heartbeat classification.

Biometric identification of cardiosynchronous waveforms utilizing person specific continuous and discrete wavelet transform features.

In this paper we explore how a Radio Frequency Impedance Interrogation (RFII) signal may be used as a biometric feature. This could allow the identification of subjects in operational and potentially hostile environments. Features extracted from the continuous and discrete wavelet decompositions of the signal are investigated for biometric identification. In the former case, the most discriminative features in the wavelet space were extracted using a Fisher ratio metric. Comparisons in the wavelet space were done using the Euclidean distance measure. In the latter case, the signal was decomposed at various levels using different wavelet bases, in order to extract both low frequency and high frequency components. Comparisons at each decomposition level were performed using the same distance measure as before. The data set used consists of four subjects, each with a 15 minute RFII recording. The various data samples for our experiments, corresponding to a single heart beat duration, were extracted from these recordings. We achieve identification rates of up to 99% using the CWT approach and rates of up to 100% using the DWT approach. While the small size of the dataset limits the interpretation of these results, further work with larger datasets is expected to develop better algorithms for subject identification.

Classification of cardiosynchronous waveforms by projection to a Legendre Polynomial sub-space.

UNLABELLED: The use of Radio Frequency Impedance Interrogation (RFII) is being investigated for use as a noninvasive hemodynamic monitoring system and in the capacity of a biometric identifier. Biometric identification of subjects by cardiosynchronous waveform generated through RFII technology could allow the identification of subjects in operational and potentially hostile environments. Here, the filtering methods for extracting a unique biometric signature from the RFII signal are examined, including the use of Cepstral analysis for dynamically estimating the filter parameters. METHODS: The projection of that signature to a Legendre Polynomial sub-space is proposed for increased class separability in a low dimensional space. Support Vector Machine (SVM) and k-Nearest Neighbor (k=3) classification are performed in the Legendre Polynomial sub-space on a small dataset. RESULTS: Both the k-Nearest Neighbor and linear SVM methods demonstrated highly successful classification accuracy, with 93-100% accuracy demonstrated by various classification methods. CONCLUSIONS: The results are highly encouraging despite the small sample size. Further analysis with a larger dataset will help to refine this process for the eventual application of RFII as a robust biometric identifier.

Customizing the training dataset to an individual for improved heartbeat recognition performance in long-term ECG signals.

This work presents an investigation of the potential benefits of customizing the analysis of long-term ECG signals, collected from individuals using wearable sensors, by incorporating small amount of data from these individuals in the training set of our classifiers. The global training dataset selected was from the MIT-BIH Arrhythmias Database. This proposal is validated on long-term ECG recordings collected via wearable technology in unsupervised environments, as well on the MIT-BIH Normal Sinus Rhythm Database. Results illustrate that heartbeat classification performance could improve significantly if short periods of data (e.g., data from the first 5-minutes of every 2 hours) from the specific individual are regularly selected and incorporated into the global training dataset for training a customized classifier.

Discrete magnitude-squared channel modeling, equalization, and detection for volume holographic storage channels.

As storage density increases, the performance of volume holographic storage channels is degraded, because intersymbol interference and noise also increase. Equalization and detection methods must be employed to mitigate the effects of intersignal interference and noise. However, the output detector array in a holographic storage system detects the intensity of the incident light's wave front, leading to loss of sign information. This sign loss precludes the applicability of conventional equalization and detection schemes. We first address channel modeling under quadratic nonlinearity and develop an efficient model named the discrete magnitude-squared channel model. We next introduce an advanced equalization method called the iterative magnitude-squared decision feedback equalization (IMSDFE), which takes the channel nonlinearity into account. The performance of IMSDFE is quantified for optical-noise-dominated channels as well as for electronic-noise-dominated channels. Results indicate that IMSDFE is a good candidate for a high-density, high-intersignal-interference volume holographic storage channel.