Search

Convolution has been the core operation of modern deep neural networks. It is well known that convolutions can be implemented in the Fourier Transform domain. In this article, we propose to use binary block Walsh–Hadamard transform (WHT) instead of the Fourier transform. We use WHT-based binary layers to replace some of the regular convolution layers in deep neural networks. We utilize both one-dimensional (1D) and 2D binary WHTs in this article. In both 1D and 2D layers, we compute the binary WHT of the input feature map and denoise the WHT domain coefficients using a nonlinearity that is obtained by combining soft-thresholding with the tanh function. After denoising, we compute the inverse WHT. We use 1D-WHT to replace the 1 × 1 convolutional layers, and 2D-WHT layers can replace the 3 × 3 convolution layers and Squeeze-and-Excite layers. 2D-WHT layers with trainable weights can be also inserted before the Global Average Pooling layers to assist the dense layers. In this way, we can reduce the number of trainable parameters significantly with a slight decrease in trainable parameters. In this article, we implement the WHT layers into MobileNet-V2, MobileNet-V3-Large, and ResNet to reduce the number of parameters significantly with negligible accuracy loss. Moreover, according to our speed test, the 2D-FWHT layer runs about 24 times as fast as the regular 3 × 3 convolution with 19.51% less RAM usage in an NVIDIA Jetson Nano experiment.

Introduction: We aim to apply deep learning methods to achieve the continuous classification of the Cervical Vertebrae Maturation (CVM) stages and to assess skeletal maturity. We propose a novel two-stage system with a parallel structure network and a sigmoid-based method to generate the continuous-valued cervical vertebrae maturity (CVCVM) parameter. Methods: A total of 1398 Cephalometric radiographs are meticulously annotated and stratified based on their respective Cervical Vertebrae Maturation (CVM) stages, with 1018 images allocated for training and validation, also the remaining 380 collected from 25 patients and labeled by two clinicians for testing. The images are further partitioned according to gender. A two-stage system is devised for the continuous estimation of CVM stages. A parallel-structure neural network called TriPodNet is trained to gauge the likelihood of each class for the maturation stage in the first part of the proposed system. The network is supplied with two different types of input, namely a radiographic X-ray image and chronological age. Probability values of individual classes are generated and mapped onto a continuous stage by two different methods, namely weighted averaging and sigmoid-based regression. The correlation of the estimated Cervical Vertebrae Maturation parameter is assessed using the Pearson correlation coefficient. In order to ascertain the validity of TriPodNet, the Permutation Importancemethod is employed to gauge the impact of each input. Results: TriPodNet is able to achieve a validation accuracy of 81.17% for female subjects and 75.96% for male subjects. During testing, the class probability values of the inputs are determined by TriPodNet, and the continuous estimation parameters are obtained by applying two distinct mapping functions. The sigmoid-based regression method produces an average correlation coefficient value of 0.910 with the first clinician and 0.944 with the second clinician for male patients, while for female patients the values were 0.910 and 0.918 respectively. Conversely, the weighted average method performs less effectively, with average correlation coefficient values of 0.913 and 0.904 for male patients with the first and second clinicians respectively. For female patients, the method produces similar results with an average correlation coefficient value of 0.901 and 0.896 with the first and second clinicians respectively. The Permutation Importance method shows that the image input and the chronological age input collaboratively contribute to the model in producing the accurate output. Conclusion: The proposed method to determine the CVM stages in a continuous stages pattern CVCVM using Convolutional Neural Network (CNN) achieved novel high correlation results compared to true labels and it is more consistent with the gradual growth changes. It is observed to be a unique way to represent skeletal maturity and assess growth.

Introduction Obstructive sleep apnea (OSA) is associated with hypertension due to intermittent hypoxia and sleep fragmentation. Due to the complex pathogenesis of hypertension, it is difficult to predict incident hypertension associated with OSA. A Machine Learning (ML) model to predict incident hypertension identified up to five years after the diagnosis of OSA by polysomnography developed. Methods Polysomnography provides time-series data on multiple physiological signals. We used the sleep heart health study (SHHS) cohort, where 4,797 participants had OSA. After excluding participants with pre-existing hypertension at baseline, the sample size was 2,652. 1,814 participants with follow-up data at 5 years were included (911/1,814, 50% with incident hypertension). In addition to clinical data (i.e. age and race), features extracted from polysomnography (heart rate variability, HRV calculated based on the electrocardiography R-R interval), electroencephalography delta power, statistical information (i.e., mean and standard deviation of signals), and heart rate periodicity functions fed to support vector machine (SVM) ML model to train and validate. The polysomnography features were calculated over the 30-second epochs identified based on respiratory events and EEG arousal and respiratory events annotation, and their corresponding parts in other signals based on sampling frequency. Technical artifacts in oxygen saturation and ECG were reconstructed with the interpolation method and removed from the signal respectively. The SVM is a robust ML method trained in an iterative fashion to find the global optimum. In comparison to the Deep Neural Network (DNN) approaches, SVMs results are interpretable. Each polysomnography signal and its corresponding features were trained on a separate SVM, followed by a fusion of the SVM results. The final results were fused by voting of individual SVM results. Results The SVM ML model thus far has achieved a test accuracy (area under the curve, AUC) of 66.06%, sensitivity 63.21%, and specificity 68.9%. Conclusion This proof-of-concept study suggests that supervised ML models, such as the SVM, may be useful in predicting incident hypertension associated with OSA. Further research is required regarding optimal input features to boost the accuracy, followed by external validation of the model in additional OSA cohorts. Support (if any) Research support 1R56HL157182, NIH/NHLBI

In this paper, we propose a computationally efficient deep learning framework to address the issue of sensitivity drift compensation for chemical sensors. The framework estimates the underlying drift signal from sensor measurements by means of a deep neural network with a multiplication-free Hadamard transform based layer. In addition, we propose an additive neural network which can be efficiently implemented in real-time on low-cost processors. The temporal additive neural network structure performs only one multiplication per “convolution” operation. Both the regular network and the additive network can have Hadamard transform based layers that implement orthogonal transforms over feature maps and perform soft-thresholding operations in the transform domain to eliminate noise. We also investigate the use of the Discrete Cosine Transform (DCT) and compare it with the Hadamard transform. We present experimental results demonstrating that the Hadamard transform outperforms the DCT.

In this article, phase of the Fourier transform (FT), which has observed to be a crucial component in image representation, is utilized for visual target tracking. The main aim of the proposed scheme is to reduce the computational complexity of cross-correlation-based matching frameworks. Normalized cross-correlation (NCC) function-based object tracker is converted to a phase minimization problem under the following assumption: In visual object tracking applications, if the frame rate is high, the moving object can be considered to have translational shifts in image domain in a small time window. Since the proposed tracking framework works in the Fourier domain, the translational shifts in the image space are converted to phase variations in the Fourier domain due to the “translational invariance” property of the FT. The proposed algorithm estimates the spatial target position based on the phase information of the target region. The proposed framework uses the $$\ell _1$$ -norm and provides a computationally efficient solution for the tracking problem. Experimental studies indicate that the proposed phase-based technique obtain comparable results with baseline tracking algorithms which are computationally more complex.

This study aims to investigate spine sounds from a perspective that would make their use for diagnostic purposes of any spinal pathology possible. People with spine problems can be determined using joint sounds collected from the involved area of the spinal columns of subjects. In our sound dataset, it is observed that a ‘click’ sound is detected in individuals who are suffering from low back pain. Recorded joint sounds are classified using automatic speech recognition algorithm. mel-frequency cepstrum coefficients (MFCC) are extracted from the sound signals as feature vectors. MFCC’s are classified using an artificial neural networks, which is currently the state-of-the-art speech recognition tool. The algorithm has a high success rate of detecting ‘click’ sounds in a given sound signal and it can perfectly identify and differentiate healthy individuals from unhealthy subjects in our data set. Spine sounds have the potential of serving as a reliable marker of the spine health.

We present a novel deep learning method for fully automated detection and classification of the Cervical Vertebrae Maturation (CVM) stages. The deep convolutional neural network consists of three parallel networks (TriPodNet) independently trained with different initialization parameters. They also have a built-in set of novel directional filters that highlight the Cervical Verte edges in X-ray images. Outputs of the three parallel networks are combined using a fully connected layer. 1018 cephalometric radiographs were labeled, divided by gender, and classified according to the CVM stages. Resulting images, using different training techniques and patches, were used to train TripodNet together with a set of tunable directional edge enhancers. Data augmentation is implemented to avoid overfitting. TripodNet achieves the state-of-the-art accuracy of 81.18\% in female patients and 75.32\% in male patients. The proposed TripodNet achieves a higher accuracy in our dataset than the Swin Transformers and the previous network models that we investigated for CVM stage estimation.

Automatic tracking of cells is a widely studied problem in various biomedical applications. Although there are numerous approaches for the video object tracking task in different contexts, the performance of these methods depends on many factors regarding the specific application they are used for. This paper presents a comparative study that specifically targets cell tracking problem and compares performance behavior of the recent algorithms. We propose a framework for the performance evaluation of the tracking algorithms and compare several state-of-the-art object tracking approaches on an extensive time-lapse inverted microscopy dataset. We report the quantitative evaluations of the algorithms based on success rate and precision performance metrics.

Introduction We aim to apply deep learning to achieve fully automated detection and classification of the Cervical Vertebrae Maturation (CVM) stages. We propose an innovative custom-designed deep Convolutional Neural Network (CNN) with a built-in set of novel directional filters that highlight the edges of the Cervical Vertebrae in X-ray images. Methods A total of 1018 Cephalometric radiographs were labeled and classified according to the Cervical Vertebrae Maturation (CVM) stages. The images were cropped to extract the cervical vertebrae using an Aggregate Channel Features (ACF) object detector. The resulting images were used to train four different Deep Learning (DL) models: our proposed CNN, MobileNetV2, ResNet101, and Xception, together with a set of tunable directional edge enhancers. When using MobileNetV2, ResNet101 and Xception, data augmentation is adopted to allow adequate network complexity while avoiding overfitting. The performance of our CNN model was compared with that of MobileNetV2, ResNet101 and Xception with and without the use of directional filters. For validation and performance assessment, k-fold cross-validation, ROC curves, and p-values were used. Results The proposed innovative model that uses a CNN preceded with a layer of tunable directional filters achieved a validation accuracy of 84.63%84.63% in CVM stage classification into five classes, exceeding the accuracy achieved with the other DL models investigated. MobileNetV2, ResNet101 and Xception used with directional filters attained accuracies of 78.54%, 74.10%, and 80.86%, respectively. The custom-designed CNN method also achieves 75.11% in six-class CVM stage classification. The effectiveness of the directional filters is reflected in the improved performance attained in the results. If the custom-designed CNN is used without the directional filters, the test accuracy decreases to 80.75%. In the Xception model without the directional filters, the testing accuracy drops slightly to 79.42% in the five-class CVM stage classification. Conclusion The proposed model of a custom-designed CNN together with the tunable Directional Filters (CNNDF) is observed to provide higher accuracy than the commonly used pre-trained network models that we investigated in the fully automated determination of the CVM stages.

In this paper, we propose a novel layer based on fast Walsh-Hadamard transform (WHT) and smooth-thresholding to replace $1\times 1$ convolution layers in deep neural networks. In the WHT domain, we denoise the transform domain coefficients using the new smooth-thresholding non-linearity, a smoothed version of the well-known soft-thresholding operator. We also introduce a family of multiplication-free operators from the basic 2$\times$2 Hadamard transform to implement $3\times 3$ depthwise separable convolution layers. Using these two types of layers, we replace the bottleneck layers in MobileNet-V2 to reduce the network's number of parameters with a slight loss in accuracy. For example, by replacing the final third bottleneck layers, we reduce the number of parameters from 2.270M to 540K. This reduces the accuracy from 95.21\% to 92.98\% on the CIFAR-10 dataset. Our approach significantly improves the speed of data processing. The fast Walsh-Hadamard transform has a computational complexity of $O(m\log_2 m)$. As a result, it is computationally more efficient than the $1\times1$ convolution layer. The fast Walsh-Hadamard layer processes a tensor in $\mathbb{R}^{10\times32\times32\times1024}$ about 2 times faster than $1\times1$ convolution layer on NVIDIA Jetson Nano computer board., The paper (v1) has been accepted to CVPR 2021 BiVision Workshop. We notice the final Conv2D is also a 1x1 convolution layer so we update the result with changing the layer in v2. In v3, we update citation 37 because its authorship changes. In v4, we propose the improved version of smooth thresholding called "weighted smooth thresholding"

In this paper, we propose a deep convolutional neural network for camera based wildfire detection. We train the neural network via transfer learning and use window based analysis strategy to increase the fire detection rate. To achieve computational efficiency, we calculate frequency response of the kernels in convolutional and dense layers and eliminate those filters with low energy impulse response. Moreover, to reduce the storage for edge devices, we compare the convolutional kernels in Fourier domain and discard similar filters using the cosine similarity measure in the frequency domain. We test the performance of the neural network with a variety of wildfire video clips and prune system performs as good as the regular network in daytime wild fire detection, and it also works well on some night wild fire video clips.

Sensor drift is a major problem in chemical sensors that requires addressing for reliable and accurate detection of chemical analytes. In this paper, we develop a causal convolutional neural network (CNN) with a Discrete Cosine Transform (DCT) layer to estimate the drift signal. In the DCT module, we apply soft-thresholding nonlinearity in the transform domain to denoise the data and obtain a sparse representation of the drift signal. The soft-threshold values are learned during training. Our results show that DCT layer-based CNNs are able to produce a slowly varying baseline drift signal. We train the CNN on synthetic data and test it on real chemical sensor data. Our results show that we can have an accurate and smooth drift estimate even when the observed sensor signal is very noisy., 5 pages, 3 figure, submitted to ICASSP 2020

We propose a co-design approach for compute-in-memory inference for deep neural networks (DNN). We use multiplication-free function approximators based on $\ell _{1}$ norm along with a co-adapted processing array and compute flow. Using the approach, we overcame many deficiencies in the current art of in-SRAM DNN processing such as the need for digital-to-analog converters (DACs) at each operating SRAM row/column, the need for high precision analog-to-digital converters (ADCs), limited support for multi-bit precision weights, and limited vector-scale parallelism. Our co-adapted implementation seamlessly extends to multi-bit precision weights, it doesn’t require DACs, and it easily extends to higher vector-scale parallelism. We also propose an SRAM-immersed successive approximation ADC (SA-ADC), where we exploit the parasitic capacitance of bit lines of SRAM array as a capacitive DAC. Since the dominant area overhead in SA-ADC comes due to its capacitive DAC, by exploiting the intrinsic parasitic of SRAM array, our approach allows low area implementation of within-SRAM SA-ADC. Our $8\times 62$ SRAM macro, which requires a 5-bit ADC, achieves ~105 tera operations per second per Watt (TOPS/W) with 8-bit input/weight processing at 45 nm CMOS. Our $8\times 30$ SRAM macro, which requires a 4-bit ADC, achieves ~84 TOPS/W. SRAM macros that require lower ADC precision are more tolerant of process variability, however, have lower TOPS/W as well. We evaluated the accuracy and performance of our proposed network for MNIST, CIFAR10, and CIFAR100 datasets. We chose a network configuration which adaptively mixes multiplication-free and regular operators. The network configurations utilize the multiplication-free operator for more than 85% operations from the total. The selected configurations are 98.6% accurate for MNIST, 90.2% for CIFAR10, and 66.9% for CIFAR100. Since most of the operations in the considered configurations are based on proposed SRAM macros, our compute-in-memory’s efficiency benefits broadly translate to the system-level.

We consider a family of vector dot products that can be implemented using sign changes and addition operations only. The dot products are energy-efficient as they avoid the multiplication operation entirely. Moreover, the dot products induce the $\ell_1$-norm, thus providing robustness to impulsive noise. First, we analytically prove that the dot products yield symmetric, positive semi-definite generalized covariance matrices, thus enabling principal component analysis (PCA). Moreover, the generalized covariance matrices can be constructed in an Energy Efficient (EEF) manner due to the multiplication-free property of the underlying vector products. We present image reconstruction examples in which our EEF PCA method result in the highest peak signal-to-noise ratios compared to the ordinary $\ell_2$-PCA and the recursive $\ell_1$-PCA., Comment: 6 pages, 3 tables and one figure

Large, open, public events, such as marathons and festivals, have always presented a unique safety challenge. These sprawling events, which can take up entire city blocks or stretch for many miles, can draw tens to hundreds of thousands of spectators and in some cases have open admission. As it is impracticable to guarantee the subjection of every event-goer to a security screening, we propose a crowd-based explosive detection system that uses a multitude of low-cost ChemFET sensors which are distributed to attendees. As the sensors offer limited accuracy, we further propose a server-based decision-making framework that utilizes a two-level feedback loop between the sensors and the server and explores spatial and temporal locality of the collected data to overcome the inherent low-accuracy of individual sensors. We thoroughly explore two distinct detection schemes, stressing their performance under a myriad of conditions, thus showing that such a crowd-based detection system comprised of low-cost and low-accuracy sensors can deliver high detection accuracy with minimal false positives.

Sudden Unexplained Death in Epilepsy (SUDEP) is a fatal threat to patients who suffer from convulsive seizures. The causes of the SUDEP are still ambiguous, and the patients who suffer from epileptic seizures may face death during sleep, likely after an unwitnessed convulsive seizure. An important step towards SUDEP prevention is reliable seizure detection during sleep that is inexpensive and unobtrusive. In this work, we developed a non-contact, nonintrusive, privacy-preserving system that can detect convulsive movements experienced by human subjects. Detection is accomplished by a combination of uncooled low-cost, low-power, low-resolution (8 × 8) IR array sensor, and a deep learning algorithm implemented with a Convolutional Neural Network (CNN). The thermopile sensor array is placed 1m from subjects who are reclining in bed. The CNN training set consists of thermal video streams from 40 healthy subjects mimicking convulsive movements or lying in bed without making convulsive movements. After training, the CNN was tested on thermal video streams not included in the training set and had a 99.2% accuracy in classifying convulsive movements and non-convulsive episodes, with no false negatives to distinguish between the occurrence and non-occurrence of convulsive movements. The performance results show that the thermopile sensor array has the potential to detect convulsive seizures while maintaining patient privacy and not requiring direct patient contact.

Electrocardiography (ECG) is a widely used tool for studying and diagnosing the heart diseases. Atrial fibrillation (AF) is an irregular and often rapid heart rate that can increase the risk of strokes, heart failure and other heart-related complications. In this study, we develop a novel and effective method to predict the potential AF risk of patients using our ECG signal dataset collected in the University of Illinois Hospital and Health Sciences System. We use a convolutional neural network (CNN) structure to process both the ECG signals and the related health data of patients. Our experimental results indicate that the model with patients’ health data can predict the AF with 79.9% accuracy), and which is better than a CNN trained without related health data 72.2% accuracy), which implies that patients’ health data play an important role in predicting AF risk. Very high sensitivity and specificity of the class of normal sinus rhythm (NSR) cases also verify that the model works well for distinguishing between NSR and ECG signals with potential AF risk.

In this paper, we propose a deep convolutional neural network for camera based wildfire detection. We train the neural network via transfer learning and use window based analysis strategy to increase the fire detection rate. To achieve computational efficiency, we calculate frequency response of the kernels in convolutional and dense layers and eliminate those filters with low energy impulse response. Moreover, to reduce the storage for edge devices, we compare the convolutional kernels in Fourier domain and discard similar filters using the cosine similarity measure in the frequency domain. We test the performance of the neural network with a variety of wildfire video clips and the pruned system performs as good as the regular network in daytime wild fire detection, and it also works well on some night wild fire video clips.

The compressed sensing (CS) framework leverages the sparsity of MR images to reconstruct from undersampled acquisitions. CS reconstructions involve one or more regularization parameters that weigh sparsity in transform domains against fidelity to acquired data. While parameter selection is critical for reconstruction quality, the optimal parameters are subject and dataset specific. Thus, commonly practiced heuristic parameter selection generalizes poorly to independent datasets. Recent studies have proposed to tune parameters by estimating the risk of removing significant image coefficients. Line searches are performed across the parameter space to identify the parameter value that minimizes this risk. Although effective, these line searches yield prolonged reconstruction times. Here, we propose a new self-tuning CS method for multi-coil multi-acquisition reconstructions. The proposed method uses computationally efficient projections onto epigraph sets of the $l_1$ and total-variation norms to simultaneously achieve parameter selection and regularization. In vivo demonstrations are provided for balanced steady-state free precession, time-of-flight, and T1-weighted imaging. The proposed method achieves nearly an order of magnitude improvement in computational efficiency over line-search methods while maintaining near-optimal parameter selection., Citation information: DOI 10.1109/TMI.2018.2885599, IEEE Transactions on Medical Imaging

In this paper, we introduce a video-based wildfire detection scheme based on a computationally efficient additive deep neural network, which we call AddNet. This AddNet is based on a multiplication-free vector operator, which performs only addition and sign manipulation operations. In this regard, we construct a dot product-like operation from the mf-operator and use it to define dense and convolutional feed-forwarding passes in AddNet. We train AddNet on images taken from forestry surveillance cameras. Our experiments show that AddNet can achieve a time-saving by 12.4% when compared to an equivalent regular convolutional neural network (CNN). Furthermore, the smoke recognition performance of AddNet is as good as regular CNNs and substantially better than binary-weight neural networks.

Date of Conference: 30 November - 3 December 2020 Conference name: 12th International Conference on Computational Collective Intelligence, ICCCI 2020 Real-time flame detection is crucial in video-based surveillance systems. We propose a vision-based method to detect flames using Deep Convolutional Generative Adversarial Neural Networks (DCGANs). Many existing supervised learning approaches using convolutional neural networks do not take temporal information into account and require a substantial amount of labeled data. To have a robust representation of sequences with and without flame, we propose a two-stage training of a DCGAN exploiting spatio-temporal flame evolution. Our training framework includes the regular training of a DCGAN with real spatio-temporal images, namely, temporal slice images, and noise vectors, and training the discriminator separately using the temporal flame images without the generator. Experimental results show that the proposed method effectively detects flame in video with negligible false-positive rates in real-time. A. Enis Çetin’s research is partially funded by NSF with grant number 1739396 and NVIDIA Corporation. B. Uğur Töreyin’s research is partially funded by TÜBİTAK 114E426, İTÜ BAP MGA-2017-40964 and MOA-2019-42321.

In this article, we present a deconvolution software based on convex sets constructed from the phase of the Fourier Transform, bounded l 2 energy and l 1 energy of a given image. The iterative deconvolution algorithm is based on the method of projections onto convex sets. Another feature of the method is that it can incorporate an approximate total variation bound called filtered variation bound on the iterative deconvolution algorithm. The main purpose of this article is to introduce the open source software called projDeconv v2.

The article reviews signal processing applications of Pyroelectric Infrared (PIR) sensors with the main focus on the respiration and pulse rate detection. Respiration and pulse rates are vital signs that are routinely measured by healthcare providers to assess a human body's basic functions. In this paper, we review non-contact human respiratory rate and heart rate (pulse) estimation methods using the low-cost, contact-free, and non-invasive Pyroelectric Infrared (PIR) sensor. The time-varying sensor signal generated by the chest and subtle head movements are used to infer the respiratory rate and the heart rate, respectively. The PIR sensor can be mounted on the head rest of a chair or placed near a bed and can continuously estimate both the respiratory rate and the heart rate of a resting person at the same time. Experimental results indicate that a positive correlation higher than 95% between the heart rate estimates and the pulse oximeter measurements is achieved in our datasets. Similarly, a correlation coefficient of 95% is also achieved between the PIR-based respiratory rate estimates and the respiratory chest movements.

In this paper, we describe a non-invasive and non-contact system of estimating resting heart rate (RHR) using a pyroelectric infrared (PIR) sensor. This infrared system monitors and records the chest motion of a subject using the analog output signal of the PIR sensor. The analog output signal represents the composite motion due to inhale-exhale process with magnitude much larger than the minute vibrations of heartbeat. Since the acceleration of the heart activity is much faster than breathing the second derivative of the PIR sensor signal monitoring the chest of the subject is used to estimate the resting heart rate. Experimental results indicate that this ambient sensor can measure resting heart rate with a chi-square significance level of α = 0.05 compared to an industry standard PPG sensor. This new system provides a low cost and an effective way to estimate the resting heart rate, which is an important biological marker.

Time-frequency techniques have difficulties in yielding efficient online algorithms for almost periodic signals. We describe a new topological method to find the period of signals that have an almost periodic waveform. Proposed method is applied to signals received from a pyro-electric infrared sensor array for the online estimation of the respiratory rate (RR) of a person. Time-varying analog signals captured from the sensors exhibit an almost periodic behavior due to repetitive nature of breathing activity. Sensor signals are transformed into two-dimensional point clouds with a technique that allows preserving the period information. Features, which represent the harmonic structures in the sensor signals, are detected by applying persistent homology and the RR is estimated based on the persistence barcode of the first Betti number. Experiments have been carried out to show that our method makes reliable estimates of the RR.

A new method to objectively measure pain using computer vision and machine learning technologies is presented. Our method seeks to capture facial expressions of pain to detect pain, especially when a patients cannot communicate pain verbally. This approach relies on using Facial muscle-based Action Units (AUs), defined by the Facial Action Coding System (FACS), that are associated with pain. It is impractical to use human FACS coding experts in clinical settings to perform this task as it is too labor-intensive and recent research has sought computer-based solutions to the problem. An effective automated system for performing the task is proposed here in which we develop an end-to-end deep learning-based Automated Facial Expression Recognition (AFER) that jointly detects the complete set of pain-related AUs. The facial video clip is processed frame by frame to estimate a vector of AU likelihood values for each frame using a deep convolutional neural network. The AU vectors are concatenated to form a table of AU values for a given video clip. Our results show significantly improved performance compared with those obtained with other known methods.

Robust and computationally efficient anomaly detection in videos is a problem in video surveillance systems. We propose a technique to increase robustness and reduce computational complexity in a Convolutional Neural Network (CNN) based anomaly detector that utilizes the optical flow information of video data. We reduce the complexity of the network by denoising the intermediate layer outputs of the CNN and by using powers-of-two weights, which replaces the computationally expensive multiplication operations with bit-shift operations. Denoising operation during inference forces small valued intermediate layer outputs to zero. The number of zeros in the network significantly increases as a result of denoising, we can implement the CNN about 10% faster than a comparable network while detecting all the anomalies in the testing set. It turns out that denoising operation also provides robustness because the contribution of small intermediate values to the final result is negligible. During training we also generate motion vector images by a Generative Adversarial Network (GAN) to improve the robustness of the overall system. We experimentally observe that the resulting system is robust to background motion.

Chemical and infra-red sensors generate distinct responses under similar conditions because of sensor drift, noise or resolution errors. In this paper, we develop novel machine learning methods for detecting and identifying VOC and Ammonia vapor from time-series data obtained by uncalibrated chemical and infrared sensors. We process time-series sensor signals using deep neural networks (DNN). Three neural network algorithms are utilized for this purpose. Additive neural networks (termed AddNet) are based on a multiplication-devoid operator and consequently exhibit energy efficiency compared to regular neural networks. The second algorithm uses generative adversarial neural networks so as to expose the classifying neural network to more realistic data points in order to help the classifier network to deliver improved generalization. Finally, we use conventional convolutional neural networks as a baseline method. Our findings indicate that using raw time-series data obtained from uncalibrated sensors and processing them using deep-learning-based methods yield better results than using hand-crafted feature parameters.

While Volatile Organic Compounds (VOC) and ammonia have a place in our daily lives, their leakage into the environment is harmful to human health. In order to prevent and detect gaseous leaks of harmful VOCs, a cyber-physical system (CPS) comprised of ordinary people or first responders is proposed. This CPS uses small, low-cost sensors coupled to smart phones or mobile devices with the necessary computation and communication capabilities. The efficacy of such a CPS hinges on its ability to address technical challenges stemming from the fact that identically produced sensors may produce different results under the same conditions due to sensor drift, noise, or resolution errors.The proposed system makes use of time-varying signals produced by sensors to detect gas leaks. Sensors sample the gas vapor level in a continuous manner and time-varying sensor data is processed using deep neural networks. One of the neural networks (NN) is an energy efficient Additive Neural Network (AddNet) which can be implemented in host devices. The second NN is the discriminator of a GAN and the third a regular convolutional NN. AddNet produces comparable VOC gas leak detection results to regular convolutional networks while reducing area requirements by two thirds.

The problem of detecting the presence of Social Anxiety Disorder (SAD) using Electroencephalography (EEG) for classification has seen limited study and is addressed with a new approach that seeks to exploit the knowledge of EEG sensor spatial configuration. Two classification models, one which ignores the configuration (model 1) and one that exploits it with different interpolation methods (model 2), are studied. Performance of these two models is examined for analyzing 34 EEG data channels each consisting of five frequency bands and further decomposed with a filter bank. The data are collected from 64 subjects consisting of healthy controls and patients with SAD. Validity of our hypothesis that model 2 will significantly outperform model 1 is borne out in the results, with accuracy $6$--$7\%$ higher for model 2 for each machine learning algorithm we investigated. Convolutional Neural Networks (CNN) were found to provide much better performance than SVM and kNNs.

Date of Conference: 12-17 May 2019 Conference Name: 44th IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 2019 Early detection of wildfire smoke in real-time is essentially important in forest surveillance and monitoring systems. We propose a vision-based method to detect smoke using Deep Convolutional Generative Adversarial Neural Networks (DC-GANs). Many existing supervised learning approaches using convolutional neural networks require substantial amount of labeled data. In order to have a robust representation of sequences with and without smoke, we propose a two-stage training of a DCGAN. Our training framework includes, the regular training of a DCGAN with real images and noise vectors, and training the discriminator separately using the smoke images without the generator. Before training the networks, the temporal evolution of smoke is also integrated with a motion-based transformation of images as a pre-processing step. Experimental results show that the proposed method effectively detects the smoke images with negligible false positive rates in real-time. The Institute of Electrical and Electronics Engineers Signal Processing Society

This paper addresses the problem of detecting epileptic seizures experienced by a human subject during sleep. Commonly used solutions to this problem mostly rely on detecting motion due to seizures using contact-based sensors or video-based sensors. We seek a low-cost, low-power alternative that can sense motion without making direct contact with the subject and provides high detection accuracy. We investigate the use of Passive InfraRed (PIR) sensors to sense human body motion caused by epileptic seizures during sleep which makes the body shake and causes the PIR sensor to generate an oscillatory output signal. This signal can be distinguished from that of ordinary motions during sleep using analysis with machine learning algorithms. The supervised hidden Markov model algorithm (HMM) and a 1-D and 2-D convolutional neural network (ConvNet) are used to classify the data set of the PIR sensor output into the occurrence of epileptic seizures, ordinary motions, or absence of motion. The method was tested on the PIR signals captured at 1 m from 33 recruited healthy subjects who, after watching seizure videos, either moved their body on a bed to simulate a seizure, ordinary motion, or lay still. The HMM algorithm attained 97.03% accuracy, while 1D-ConvNet and 2D-ConvNet attained an accuracy of 96.97% and 98.98%, respectively. All simulated seizures were successfully detected, with errors occurring only in distinguishing between ordinary motion and no motion, thereby demonstrating the potential for using PIR sensors in the epileptic seizure detection.