Implementation of Sound Direction Detection and Mixed Source Separation in Embedded Systems.

In recent years, embedded system technologies and products for sensor networks and wearable devices used for monitoring people's activities and health have become the focus of the global IT industry. In order to enhance the speech recognition capabilities of wearable devices, this article discusses the implementation of audio positioning and enhancement in embedded systems using embedded algorithms for direction detection and mixed source separation. The two algorithms are implemented using different embedded systems: direction detection developed using TI TMS320C6713 DSK and mixed source separation developed using Raspberry Pi 2. For mixed source separation, in the first experiment, the average signal-to-interference ratio (SIR) at 1 m and 2 m distances was 16.72 and 15.76, respectively. In the second experiment, when evaluated using speech recognition, the algorithm improved speech recognition accuracy to 95%.

Progressively Unsupervised Generative Attentional Networks with Adaptive Layer-Instance Normalization for Image-to-Image Translation.

Unsupervised image-to-image translation has received considerable attention due to the recent remarkable advancements in generative adversarial networks (GANs). In image-to-image translation, state-of-the-art methods use unpaired image data to learn mappings between the source and target domains. However, despite their promising results, existing approaches often fail in challenging conditions, particularly when images have various target instances and a translation task involves significant transitions in shape and visual artifacts when translating low-level information rather than high-level semantics. To tackle the problem, we propose a novel framework called Progressive Unsupervised Generative Attentional Networks with Adaptive Layer-Instance Normalization (PRO-U-GAT-IT) for the unsupervised image-to-image translation task. In contrast to existing attention-based models that fail to handle geometric transitions between the source and target domains, our model can translate images requiring extensive and holistic changes in shape. Experimental results show the superiority of the proposed approach compared to the existing state-of-the-art models on different datasets.

Electrocardiogram Heartbeat Classification for Arrhythmias and Myocardial Infarction.

An electrocardiogram (ECG) is a basic and quick test for evaluating cardiac disorders and is crucial for remote patient monitoring equipment. An accurate ECG signal classification is critical for real-time measurement, analysis, archiving, and transmission of clinical data. Numerous studies have focused on accurate heartbeat classification, and deep neural networks have been suggested for better accuracy and simplicity. We investigated a new model for ECG heartbeat classification and found that it surpasses state-of-the-art models, achieving remarkable accuracy scores of 98.5% on the Physionet MIT-BIH dataset and 98.28% on the PTB database. Furthermore, our model achieves an impressive F1-score of approximately 86.71%, outperforming other models, such as MINA, CRNN, and EXpertRF on the PhysioNet Challenge 2017 dataset.

Heuristic Attention Representation Learning for Self-Supervised Pretraining.

Recently, self-supervised learning methods have been shown to be very powerful and efficient for yielding robust representation learning by maximizing the similarity across different augmented views in embedding vector space. However, the main challenge is generating different views with random cropping; the semantic feature might exist differently across different views leading to inappropriately maximizing similarity objective. We tackle this problem by introducing Heuristic Attention Representation Learning (HARL). This self-supervised framework relies on the joint embedding architecture in which the two neural networks are trained to produce similar embedding for different augmented views of the same image. HARL framework adopts prior visual object-level attention by generating a heuristic mask proposal for each training image and maximizes the abstract object-level embedding on vector space instead of whole image representation from previous works. As a result, HARL extracts the quality semantic representation from each training sample and outperforms existing self-supervised baselines on several downstream tasks. In addition, we provide efficient techniques based on conventional computer vision and deep learning methods for generating heuristic mask proposals on natural image datasets. Our HARL achieves +1.3% advancement in the ImageNet semi-supervised learning benchmark and +0.9% improvement in AP50 of the COCO object detection task over the previous state-of-the-art method BYOL. Our code implementation is available for both TensorFlow and PyTorch frameworks.

Embedding Biometric Information in Interpolated Medical Images with a Reversible and Adaptive Strategy.

How to hide messages in digital images so that messages cannot be discovered and tampered with is a compelling topic in the research area of cybersecurity. The interpolation-based reversible data hiding (RDH) scheme is especially useful for the application of medical image management. The biometric information of patients acquired by biosensors is embedded into an interpolated medical image for the purpose of authentication. The proposed scheme classifies pixel blocks into complex and smooth ones according to each block's dynamic range of pixel values. For a complex block, the minimum-neighbor (MN) interpolation followed by DIM embedding is applied, where DIM denotes the difference between the block's interpolated pixel values and the maximum pixel values. For a smooth block, the block mean (BM) interpolation is followed by a prediction error histogram (PEH) embedding and a difference expansion (DE) embedding is applied. Compared with previous methods, this adaptive strategy ensures low distortion due to embedding for smooth blocks while it provides a good payload for complex blocks. Our scheme is suitable for both medical and general images. Experimental results confirm the effectiveness of the proposed scheme. Performance comparisons with state-of-the-art schemes are also given. The peak signal to noise ratio (PSNR) of the proposed scheme is 10.32 dB higher than the relevant works in the best case.

A Novel Method for Baroreflex Sensitivity Estimation Using Modulated Gaussian Filter.

The evaluation of baroreflex sensitivity (BRS) has proven to be critical for medical applications. The use of α indices by spectral methods has been the most popular approach to BRS estimation. Recently, an algorithm termed Gaussian average filtering decomposition (GAFD) has been proposed to serve the same purpose. GAFD adopts a three-layer tree structure similar to wavelet decomposition but is only constructed by Gaussian windows in different cutoff frequency. Its computation is more efficient than that of conventional spectral methods, and there is no need to specify any parameter. This research presents a novel approach, referred to as modulated Gaussian filter (modGauss) for BRS estimation. It has a more simplified structure than GAFD using only two bandpass filters of dedicated passbands, so that the three-level structure in GAFD is avoided. This strategy makes modGauss more efficient than GAFD in computation, while the advantages of GAFD are preserved. Both GAFD and modGauss are conducted extensively in the time domain, yet can achieve similar results to conventional spectral methods. In computational simulations, the EuroBavar dataset was used to assess the performance of the novel algorithm. The BRS values were calculated by four other methods (three spectral approaches and GAFD) for performance comparison. From a comparison using the Wilcoxon rank sum test, it was found that there was no statistically significant dissimilarity; instead, very good agreement using the intraclass correlation coefficient (ICC) was observed. The modGauss algorithm was also found to be the fastest in computation time and suitable for the long-term estimation of BRS. The novel algorithm, as described in this report, can be applied in medical equipment for real-time estimation of BRS in clinical settings.

Self-Supervised Learning Framework toward State-of-the-Art Iris Image Segmentation.

Iris segmentation plays a pivotal role in the iris recognition system. The deep learning technique developed in recent years has gradually been applied to iris recognition techniques. As we all know, applying deep learning techniques requires a large number of data sets with high-quality manual labels. The larger the amount of data, the better the algorithm performs. In this paper, we propose a self-supervised framework utilizing the pix2pix conditional adversarial network for generating unlimited diversified iris images. Then, the generated iris images are used to train the iris segmentation network to achieve state-of-the-art performance. We also propose an algorithm to generate iris masks based on 11 tunable parameters, which can be generated randomly. Such a framework can generate an unlimited amount of photo-realistic training data for down-stream tasks. Experimental results demonstrate that the proposed framework achieved promising results in all commonly used metrics. The proposed framework can be easily generalized to any object segmentation task with a simple fine-tuning of the mask generation algorithm.

State-of-the-Art Capability of Convolutional Neural Networks to Distinguish the Signal in the Ionosphere.

Recovering and distinguishing different ionospheric layers and signals usually requires slow and complicated procedures. In this work, we construct and train five convolutional neural network (CNN) models: DeepLab, fully convolutional DenseNet24 (FC-DenseNet24), deep watershed transform (DWT), Mask R-CNN, and spatial attention-UNet (SA-UNet) for the recovery of ionograms. The performance of the models is evaluated by intersection over union (IoU). We collect and manually label 6131 ionograms, which are acquired from a low-latitude ionosonde in Taiwan. These ionograms are contaminated by strong quasi-static noise, with an average signal-to-noise ratio (SNR) equal to 1.4. Applying the five models to these noisy ionograms, we show that the models can recover useful signals with IoU > 0.6. The highest accuracy is achieved by SA-UNet. For signals with less than 15% of samples in the data set, they can be recovered by Mask R-CNN to some degree (IoU > 0.2). In addition to the number of samples, we identify and examine the effects of three factors: (1) SNR, (2) shape of signal, (3) overlapping of signals on the recovery accuracy of different models. Our results indicate that FC-DenseNet24, DWT, Mask R-CNN and SA-UNet are capable of identifying signals from very noisy ionograms (SNR < 1.4), overlapping signals can be well identified by DWT, Mask R-CNN and SA-UNet, and that more elongated signals are better identified by all models.

Super-Resolution Generative Adversarial Network Based on the Dual Dimension Attention Mechanism for Biometric Image Super-Resolution.

There exist many types of intelligent security sensors in the environment of the Internet of Things (IoT) and cloud computing. Among them, the sensor for biometrics is one of the most important types. Biometric sensors capture the physiological or behavioral features of a person, which can be further processed with cloud computing to verify or identify the user. However, a low-resolution (LR) biometrics image causes the loss of feature details and reduces the recognition rate hugely. Moreover, the lack of resolution negatively affects the performance of image-based biometric technology. From a practical perspective, most of the IoT devices suffer from hardware constraints and the low-cost equipment may not be able to meet various requirements, particularly for image resolution, because it asks for additional storage to store high-resolution (HR) images, and a high bandwidth to transmit the HR image. Therefore, how to achieve high accuracy for the biometric system without using expensive and high-cost image sensors is an interesting and valuable issue in the field of intelligent security sensors. In this paper, we proposed DDA-SRGAN, which is a generative adversarial network (GAN)-based super-resolution (SR) framework using the dual-dimension attention mechanism. The proposed model can be trained to discover the regions of interest (ROI) automatically in the LR images without any given prior knowledge. The experiments were performed on the CASIA-Thousand-v4 and the CelebA datasets. The experimental results show that the proposed method is able to learn the details of features in crucial regions and achieve better performance in most cases.

Mask Attention-SRGAN for Mobile Sensing Networks.

Biometrics has been shown to be an effective solution for the identity recognition problem, and iris recognition, as well as face recognition, are accurate biometric modalities, among others. The higher resolution inside the crucial region reveals details of the physiological characteristics which provides discriminative information to achieve extremely high recognition rate. Due to the growing needs for the IoT device in various applications, the image sensor is gradually integrated in the IoT device to decrease the cost, and low-cost image sensors may be preferable than high-cost ones. However, low-cost image sensors may not satisfy the minimum requirement of the resolution, which definitely leads to the decrease of the recognition accuracy. Therefore, how to maintain high accuracy for biometric systems without using expensive high-cost image sensors in mobile sensing networks becomes an interesting and important issue. In this paper, we proposed MA-SRGAN, a single image super-resolution (SISR) algorithm, based on the mask-attention mechanism used in Generative Adversarial Network (GAN). We modified the latest state-of-the-art (nESRGAN+) in the GAN-based SR model by adding an extra part of a discriminator with an additional loss term to force the GAN to pay more attention within the region of interest (ROI). The experiments were performed on the CASIA-Thousand-v4 dataset and the Celeb Attribute dataset. The experimental results show that the proposed method successfully learns the details of features inside the crucial region by enhancing the recognition accuracies after image super-resolution (SR).

Continuous Blood Pressure Estimation Using Exclusively Photopletysmography by LSTM-Based Signal-to-Signal Translation.

Monitoring continuous BP signal is an important issue, because blood pressure (BP) varies over days, minutes, or even seconds for short-term cases. Most of photoplethysmography (PPG)-based BP estimation methods are susceptible to noise and only provides systolic blood pressure (SBP) and diastolic blood pressure (DBP) prediction. Here, instead of estimating a discrete value, we focus on different perspectives to estimate the whole waveform of BP. We propose a novel deep learning model to learn how to perform signal-to-signal translation from PPG to arterial blood pressure (ABP). Furthermore, using a raw PPG signal only as the input, the output of the proposed model is a continuous ABP signal. Based on the translated ABP signal, we extract the SBP and DBP values accordingly to ease the comparative evaluation. Our prediction results achieve average absolute error under 5 mmHg, with 70% confidence for SBP and 95% confidence for DBP without complex feature engineering. These results fulfill the standard from Association for the Advancement of Medical Instrumentation (AAMI) and the British Hypertension Society (BHS) with grade A. From the results, we believe that our model is applicable and potentially boosts the accuracy of an effective signal-to-signal continuous blood pressure estimation.

Robust Iris Segmentation Algorithm in Non-Cooperative Environments Using Interleaved Residual U-Net.

Iris segmentation plays an important and significant role in the iris recognition system. The prerequisite for accurate iris recognition is the correctness of iris segmentation. However, the efficiency and robustness of traditional iris segmentation methods are severely challenged in a non-cooperative environment because of unfavorable factors, for instance, occlusion, blur, low resolution, off-axis, motion, and specular reflections. All of the above factors seriously reduce the accuracy of iris segmentation. In this paper, we present a novel iris segmentation algorithm that localizes the outer and inner boundaries of the iris image. We propose a neural network model called "Interleaved Residual U-Net" (IRUNet) for semantic segmentation and iris mask synthesis. The K-means clustering is applied to select saliency points set in order to recover the outer boundary of the iris, whereas the inner border is recovered by selecting another set of saliency points on the inner side of the mask. Experimental results demonstrate that the proposed iris segmentation algorithm can achieve the mean IOU value of 98.9% and 97.7% for inner and outer boundary estimation, respectively, which outperforms the existing approaches on the challenging CASIA-Iris-Thousand database.

Recovery of Ionospheric Signals Using Fully Convolutional DenseNet and Its Challenges.

The technique of active ionospheric sounding by ionosondes requires sophisticated methods for the recovery of experimental data on ionograms. In this work, we applied an advanced algorithm of deep learning for the identification and classification of signals from different ionospheric layers. We collected a dataset of 6131 manually labeled ionograms acquired from low-latitude ionosondes in Taiwan. In the ionograms, we distinguished 11 different classes of the signals according to their ionospheric layers. We developed an artificial neural network, FC-DenseNet24, based on the FC-DenseNet convolutional neural network. We also developed a double-filtering algorithm to reduce incorrectly classified signals. That made it possible to successfully recover the sporadic E layer and the F2 layer from highly noise-contaminated ionograms whose mean signal-to-noise ratio was low, SNR = 1.43. The Intersection over Union (IoU) of the recovery of these two signal classes was greater than 0.6, which was higher than the previous models reported. We also identified three factors that can lower the recovery accuracy: (1) smaller statistics of samples; (2) mixing and overlapping of different signals; (3) the compact shape of signals.

Real-Time Cuffless Continuous Blood Pressure Estimation Using Deep Learning Model.

Blood pressure monitoring is one avenue to monitor people's health conditions. Early detection of abnormal blood pressure can help patients to get early treatment and reduce mortality associated with cardiovascular diseases. Therefore, it is very valuable to have a mechanism to perform real-time monitoring for blood pressure changes in patients. In this paper, we propose deep learning regression models using an electrocardiogram (ECG) and photoplethysmogram (PPG) for the real-time estimation of systolic blood pressure (SBP) and diastolic blood pressure (DBP) values. We use a bidirectional layer of long short-term memory (LSTM) as the first layer and add a residual connection inside each of the following layers of the LSTMs. We also perform experiments to compare the performance between the traditional machine learning methods, another existing deep learning model, and the proposed deep learning models using the dataset of Physionet's multiparameter intelligent monitoring in intensive care II (MIMIC II) as the source of ECG and PPG signals as well as the arterial blood pressure (ABP) signal. The results show that the proposed model outperforms the existing methods and is able to achieve accurate estimation which is promising in order to be applied in clinical practice effectively.

Generalized Deep Neural Network Model for Cuffless Blood Pressure Estimation with Photoplethysmogram Signal Only.

Due to the growing public awareness of cardiovascular disease (CVD), blood pressure (BP) estimation models have been developed based on physiological parameters extracted from both electrocardiograms (ECGs) and photoplethysmograms (PPGs). Still, in order to enhance the usability as well as reduce the sensor cost, researchers endeavor to establish a generalized BP estimation model using only PPG signals. In this paper, we propose a deep neural network model capable of extracting 32 features exclusively from PPG signals for BP estimation. The effectiveness and accuracy of our proposed model was evaluated by the root mean square error (RMSE), mean absolute error (MAE), the Association for the Advancement of Medical Instrumentation (AAMI) standard and the British Hypertension Society (BHS) standard. Experimental results showed that the RMSEs in systolic blood pressure (SBP) and diastolic blood pressure (DBP) are 4.643 mmHg and 3.307 mmHg, respectively, across 9000 subjects, with 80.63% of absolute errors among estimated SBP records lower than 5 mmHg and 90.19% of absolute errors among estimated DBP records lower than 5 mmHg. We demonstrated that our proposed model has remarkably high accuracy on the largest BP database found in the literature, which shows its effectiveness compared to some prior works.

A Novel Anti-Spoofing Solution for Iris Recognition Toward Cosmetic Contact Lens Attack Using Spectral ICA Analysis.

In this study, we maneuvered a dual-band spectral imaging system to capture an iridal image from a cosmetic-contact-lens-wearing subject. By using the independent component analysis to separate individual spectral primitives, we successfully distinguished the natural iris texture from the cosmetic contact lens (CCL) pattern, and restored the genuine iris patterns from the CCL-polluted image. Based on a database containing 200 test image pairs from 20 CCL-wearing subjects as the proof of concept, the recognition accuracy (False Rejection Rate: FRR) was improved from FRR = 10.52% to FRR = 0.57% with the proposed ICA anti-spoofing scheme.

Test of the Practicality and Feasibility of EDoF-Empowered Image Sensors for Long-Range Biometrics.

For many practical applications of image sensors, how to extend the depth-of-field (DoF) is an important research topic; if successfully implemented, it could be beneficial in various applications, from photography to biometrics. In this work, we want to examine the feasibility and practicability of a well-known "extended DoF" (EDoF) technique, or "wavefront coding," by building real-time long-range iris recognition and performing large-scale iris recognition. The key to the success of long-range iris recognition includes long DoF and image quality invariance toward various object distance, which is strict and harsh enough to test the practicality and feasibility of EDoF-empowered image sensors. Besides image sensor modification, we also explored the possibility of varying enrollment/testing pairs. With 512 iris images from 32 Asian people as the database, 400-mm focal length and F/6.3 optics over 3 m working distance, our results prove that a sophisticated coding design scheme plus homogeneous enrollment/testing setups can effectively overcome the blurring caused by phase modulation and omit Wiener-based restoration. In our experiments, which are based on 3328 iris images in total, the EDoF factor can achieve a result 3.71 times better than the original system without a loss of recognition accuracy.

Anti-Aliasing Attention U-net Model for Skin Lesion Segmentation.

The need for a lightweight and reliable segmentation algorithm is critical in various biomedical image-prediction applications. However, the limited quantity of data presents a significant challenge for image segmentation. Additionally, low image quality negatively impacts the efficiency of segmentation, and previous deep learning models for image segmentation require large parameters with hundreds of millions of computations, resulting in high costs and processing times. In this study, we introduce a new lightweight segmentation model, the mobile anti-aliasing attention u-net model (MAAU), which features both encoder and decoder paths. The encoder incorporates an anti-aliasing layer and convolutional blocks to reduce the spatial resolution of input images while avoiding shift equivariance. The decoder uses an attention block and decoder module to capture prominent features in each channel. To address data-related problems, we implemented data augmentation methods such as flip, rotation, shear, translate, and color distortions, which enhanced segmentation efficiency in the international Skin Image Collaboration (ISIC) 2018 and PH2 datasets. Our experimental results demonstrated that our approach had fewer parameters, only 4.2 million, while it outperformed various state-of-the-art segmentation methods.

Correlations of Neck/Shoulder Perfusion Characteristics and Pain Symptoms of the Female Office Workers with Sedentary Lifestyle.

AIM: Modern office workers are often impacted by chronic neck/shoulder pain. Most of the previous studies which investigated the relationship of the occupational factors and musculoskeletal symptoms had adopted questionnaire survey. In this study the microcirculatory characteristics and perceived symptoms in neck/shoulder region were compared among office workers with sedentary lifestyle. METHODS: Thirty-seven female office workers were recruited in this study. Microcirculatory flow in neck/shoulder region characterized by the mean blood flow (MMBF value), pulsatile blood flow (PMBF value), and the PMBF/MMBF ratio (perfusion pulsatility, PP) were investigated using Laser Doppler Flowmetry (LDF). A Chinese version of the Standardized Nordic Musculoskeletal Questionnaire (NMQ) were also administered to collect the information of perceived neck/shoulder symptoms. Correlations between the perfusion characteristics and the individual/occupational factors were analyzed using the Spearman test. The difference of the MMBF values between the low-pain group (pain level≤2) and the high-pain group (pain level>2) were compared using the Mann-Whitney U test. RESULTS: There were 81% participants reported neck or shoulder pain symptoms. The duration of shoulder pain was significantly correlated with the workers' age and the duration of employment (p<0.01) (n = 37). While both the MMBF and PMBF values in shoulder region were significantly reduced with the workers' age and the duration of employment (p<0.05) (n = 27). And there was a 54% reduction in the MMBF value of the workers from age of 23 to 47. And the MMBF value of the high-pain group (n = 15) was significantly lower than the value of the low-pain group (n = 15) (p<0.05). The duration of shoulder pain showed a moderately negative correlation with PMBF values (n = 19). Besides, the PP value was moderately correlated with shoulder pain level attributed by the rapid reduction of MMBF values (p = 0.07). CONCLUSION: In this study, the LDF method was used for the first time in the workplace in Taiwan. It was demonstrated that the MMBF in shoulder region were affected by aging effect and towards lower value at higher pain level. Impaired microcirculation caused by age effect, when coupled with sedentary lifestyle, was found to be more likely to evoke ischemia shoulder pain. Further studies are needed to assess current indicator, PP value, and the underlying mechanism of pain caused by sedentary lifestyle.

Extending the Capture Volume of an Iris Recognition System Using Wavefront Coding and Super-Resolution.

Iris recognition has gained increasing popularity over the last few decades; however, the stand-off distance in a conventional iris recognition system is too short, which limits its application. In this paper, we propose a novel hardware-software hybrid method to increase the stand-off distance in an iris recognition system. When designing the system hardware, we use an optimized wavefront coding technique to extend the depth of field. To compensate for the blurring of the image caused by wavefront coding, on the software side, the proposed system uses a local patch-based super-resolution method to restore the blurred image to its clear version. The collaborative effect of the new hardware design and software post-processing showed great potential in our experiment. The experimental results showed that such improvement cannot be achieved by using a hardware-or software-only design. The proposed system can increase the capture volume of a conventional iris recognition system by three times and maintain the system's high recognition rate.