Search

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%.

Background: There is a lack of published research on the impact of the first wave of the COVID-19 pandemic in Taiwan. We investigated the mortality risk factors among critically ill patients with COVID-19 in Taiwan during the initial wave. Furthermore, we aim to develop a novel AI mortality prediction model using chest X-ray (CXR) alone. Method: We retrospectively reviewed the medical records of patients with COVID-19 at Taipei Tzu Chi Hospital from May 15 to July 15 2021. We enrolled adult patients who received invasive mechanical ventilation. The CXR images of each enrolled patient were divided into 4 categories (1st, pre-ETT, ETT, and WORST). To establish a prediction model, we used the MobilenetV3-Small model with “Imagenet” pretrained weights, followed by high Dropout regularization layers. We trained the model with these data with Five-Fold Cross-Validation to evaluate model performance. Result: A total of 64 patients were enrolled. The overall mortality rate was 45%. The median time from symptom onset to intubation was 8 days. Vasopressor use and a higher BRIXIA score on the WORST CXR were associated with an increased risk of mortality. The areas under the curve of the 1st, pre-ETT, ETT, and WORST CXRs by the AI model were 0.87, 0.92, 0.96, and 0.93 respectively. Conclusion: The mortality rate of COVID-19 patients who receive invasive mechanical ventilation was high. Septic shock and high BRIXIA score were clinical predictors of mortality. The novel AI mortality prediction model using CXR alone exhibited a high performance.

Speech separation has been employed in important applications such as automatic speech, paralinguistics, speech recognition, hearing aids, and human-machine interactions. In recent years, deep neural networks have been widely used for speech and music separation. Some of these breakthrough successful models based on embedding vectors have been proposed, such as deep clustering. In this paper, we propose a node encoder Squash-norm deep clustering (ESDC) as an enhanced discriminative learning framework by combining node encoder, Squash-norm, and deep clustering (DC). First, a node encoder is used to create intermediate features. Node encoders are developed through a matrix factorization-based learning method for graph representations. It creates distinguishable intermediate features that play an important role in improving performance. These discriminated intermediate features are then used as input features for the separation block. The decoder block finally constructs the estimation mask through the clustering method and reconstructs the estimated signal for each source. In particular, we apply a normalization function, Squash-norm, to the input and output vectors to enhance the distinction between high-dimensional embedding vectors. This nonlinear function amplifies the differences in the input vectors, resulting in highly unique features, which are scalar products of the vectors. Similar to the input vector, Squash-norm also enhances the discrimination of the output vector, thereby enhancing the ability to construct an estimated mask by clustering the output vector. Overall, the proposed ESDC achieves 1.27–2.09 dB SDR, 1.28–2.21 dB SDRi, and 1.3–2.44 dB SI-SNRi gain compared to the DC baseline separation performance across genders on the TSP and TIMIT datasets. With the same gender, our proposed ESDC achieves 1.14–2.71 dB SDR, 0.99–2.74 dB SDRi, and 0.62–2.86 dB SI-SNRi gain compared with the DC baseline on the TIMIT dataset. In all cases, the proposed ESDC model consistently maintains STOI and PESQ higher than the DC baselines on the TSP and TIMIT datasets.

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.

Knowledge distillation, which is a process of transferring complex knowledge learned by a heavy network, i.e., a teacher, to a lightweight network, i.e., a student, has emerged as an effective technique for compressing neural networks. To reduce the necessity of training a large teacher network, this paper leverages the recent self-knowledge distillation approach to train a student network progressively by distilling its own knowledge without a pre-trained teacher network. Far from the existing self-knowledge distillation methods, which mainly focus on still images, our proposed Teaching Yourself is a self-knowledge distillation technique that targets at videos for human action recognition. Our proposed Teaching Yourself is not only designed as an effective lightweight network but also a high generalization capability model. In our approach, the network is able to update itself using the best past model, termed the preceding model, which is then utilized to guide the training process to update the present model. Inspired by consistency training in state-of-the-art semi-supervised learning methods, we also introduce an effective augmentation strategy to increase data diversity and improve network generalization and consistent predictions for our proposed Teaching Yourself approach. Our benchmark has been conducted on both the 3D Resnet-18 and 3D ResNet-50 backbone networks and evaluated on various standard datasets such as UCF101, HMDB51, and Kinetics400 datasets. The experimental results have shown that our teaching yourself method significantly improves the action recognition performance in terms of accuracy compared to existing supervised learning and knowledge distillation methods. We also have conducted an expensive ablation study to demonstrate that our approach mitigates overconfident predictions on dark knowledge and generates more consistent predictions in input variations of the same data point. The code is available at https://github.com/vdquang1991/Self-KD.

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.

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.

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.

Accurately segmented nuclei are important, not only for cancer classification, but also for predicting treatment effectiveness and other biomedical applications. However, the diversity of cell types, various external factors, and illumination conditions make nucleus segmentation a challenging task. In this work, we present a new deep learning-based method for cell nucleus segmentation. The proposed convolutional blur attention (CBA) network consists of downsampling and upsampling procedures. A blur attention module and a blur pooling operation are used to retain the feature salience and avoid noise generation in the downsampling procedure. A pyramid blur pooling (PBP) module is proposed to capture the multi-scale information in the upsampling procedure. The superiority of the proposed method has been compared with a few prior segmentation models, namely U-Net, ENet, SegNet, LinkNet, and Mask RCNN on the 2018 Data Science Bowl (DSB) challenge dataset and the multi-organ nucleus segmentation (MoNuSeg) at MICCAI 2018. The Dice similarity coefficient and some evaluation matrices, such as F1 score, recall, precision, and average Jaccard index (AJI) were used to evaluate the segmentation efficiency of these models. Overall, the proposal method in this paper has the best performance, the AJI indicator on the DSB dataset and MoNuSeg is 0.8429, 0.7985, respectively.

Abstract In this paper, a dimensionality reduction method applied on facial expression recognition is investigated. An unsupervised learning framework, projective complex matrix factorization (proCMF), is introduced to project high-dimensional input facial images into a lower dimension subspace. The proCMF model is related to both the conventional projective nonnegative matrix factorization (proNMF) and the cosine dissimilarity metric in the simple manner by transforming real data into the complex domain. A projective matrix is then found through solving an unconstraint complex optimization problem. The gradient descent method was utilized to optimize a complex cost function. Extensive experiments carried on the extended Cohn-Kanade and the JAFFE databases show that the proposed proCMF model provides even better performance than state-of-the-art methods for facial expression recognition.

For music emotion detection, this paper presents a music emotion verification system based on hierarchical sparse kernel machines. With the proposed system, we intend to verify if a music clip possesses happiness emotion or not. There are two levels in the hierarchical sparse kernel machines. In the first level, a set of acoustical features are extracted, and principle component analysis (PCA) is implemented to reduce the dimension. The acoustical features are utilized to generate the first-level decision vector, which is a vector with each element being a significant value of an emotion. The significant values of eight main emotional classes are utilized in this paper. To calculate the significant value of an emotion, we construct its 2-class SVM with calm emotion as the global (non-target) side of the SVM. The probability distributions of the adopted acoustical features are calculated and the probability product kernel is applied in the first-level SVMs to obtain first-level decision vector feature. In the second level of the hierarchical system, we merely construct a 2-class relevance vector machine (RVM) with happiness as the target side and other emotions as the background side of the RVM. The first-level decision vector is used as the feature with conventional radial basis function kernel. The happiness verification threshold is built on the probability value. In the experimental results, the detection error tradeoff (DET) curve shows that the proposed system has a good performance on verifying if a music clip reveals happiness emotion.

The investigations of emotional speech identification can be divided into two main parts, features and classifiers. In this paper, how to extract an effective speech feature set for the emotional speech identification is addressed. In our speech feature set, we use not only statistical analysis of frame-based acoustical features, but also the approximated speech feature contours, which are obtained by extracting extremely low frequency components to speech feature contours. Furthermore, principal component analysis (PCA) is applied to the approximated speech feature contours so that an efficient representation of approximated contours can be derived. The proposed speech feature set is fed into support vector machines (SVMs) to perform multiclass emotion identification. The experimental results demonstrate the performance of the proposed system with 82.26% identification rate.

An entropy-based criterion is proposed to characterize the pattern and intensity of object motion in a video sequence as a function of time. By applying a homoscedastic error model-based time series change point detection algorithm to this motion entropy curve, one is able to segment the corresponding video sequence into individual sections, each consisting of a semantically relevant event. The proposed method is tested on six hours of sports videos including basketball, soccer, and tennis. Excellent experimental results are observed.