Search

Safeguarding personal identity in various surveillance videos, dashcams, and on-street videos is crucial. One way to do this is to detect faces and blur them, but a better solution is to replace them with non-existent ones to maintain the naturalness of the videos. While face swapping methods have already been used in the media industry with high-quality faces, it is challenging to apply them for identity protection to faces in-the-wild where faces are often occluded and of low-resolution. Therefore, we propose a new framework for face swapping specifically designed to work with face images taken in real-world scenarios, making it useful as a privacy protection method. To tackle the issue of low-resolution images, we introduce a Cross-Resolution Contrastive Loss (CRCL) technique, which allows our neural network model to be trained using triplets of varying resolutions. This enables the model to learn and use identity information across different resolutions, thereby improving its accuracy. We also propose a plug-and-play framework that can be easily applied to existing face swapping models to handle occlusions. By explicit swapping of facial features and filling of occluded regions, our framework provides a more seamless blend. To demonstrate the effectiveness of our method in handling faces in-the-wild, we create an occluded VGGFace2 dataset consisting of face images augmented with various facial masks and hand occlusions. Through quantitative and qualitative assessments on this dataset, our proposed method demonstrates robust performance under low-resolution or occluded scenarios. Significant improvements are made in the quality of swapped faces while preserving their identity and attributes, highlighting the effectiveness of our framework in advancing face swapping as a reliable privacy protection measure.

In this paper, we consider a distributed antenna system (DAS) with polarized antennas for the industrial internet-of-things, where distributed sensor nodes (SNs) with a single antenna are served by the closest distributed antenna (DA) port having multiple antennas. By using stochastic geometry, the interference in the polarized DAS is analyzed and the channel outage probability (OP) of the polarized DAS is also derived in a closed form. From the analytic results, the optimal vertically/horizontally polarized node densities are derived to minimize the channel OP. Furthermore, by investigating the effects of node density and polarization on the channel OP, we also develop a dual-polarized DAS, in which the effective density of co-polarized DA ports is the same with the density of all DA ports and, simultaneously, the interference from the active cross-polarized DA ports is effectively mitigated. By analyzing the channel OP for the dual-polarized DAS, the optimal node densities to minimize the overall channel OP and the effect of network design parameters on the channel OP is also investigated, which gives us a useful insight into the design of the polarized DAS.

Early research focused on developing effective algorithms for Ventricular fibrillation (VF) detection; while most of the evaluations have been conducted offline with prefiltered data sets, practical application requires these tests to be performed in real time. Because there are many factors that may impact detection effectiveness, it is important to understand the impact of factors that improve detection accuracy. In this study, we developed an integrated simulated environment using IAR Embedded Workbench software to build an embedded system using a MSP430 microcontroller and Visual studio tool for S/W build; we then used this system to conduct real-time experiments for evaluating five lightweight VF detection algorithms and to examine factors that may impact their performance in terms of sensitivity, specificity, positive-predictivity, accuracy and computational time. The results were cross-validated using a prototype of a wearable Electrocardiogram (ECG) system developed by this study. The study showed that 1) the chosen detection algorithm, data filtering, and window size all have a significant impact on the performance of real-time VF detection; among these, the detection algorithm had the greatest impact so it must be carefully selected; 2) it is important to select the proper threshold value that affects tradeoffs in performance metrics. Among the five algorithms that this study evaluated, the Time Delay (TD) algorithm outperformed the others independent of window size or filtering method. Considering the tradeoff between robustness and efficiency, TD is preferable because detection accuracy and robustness are more critical.

In this paper, a compact and broadband electrostatic discharge (ESD) protection input/output (I/O) pad for high-speed interfaces is designed and miniaturized using a pad-stacked inductor in a 28-nm CMOS technology. A $\pi $ -diode with a single inductor is adopted to compensate for the parasitic capacitance and extend the bandwidth of ESD-protected I/O pad. To minimize the increase of the chip area by the inductor, the pad-stacked inductor is proposed, where the inductor is placed below I/O pad. The eddy current, then, is induced in the I/O pad by time-varying magnetic field by the inductor, degrading the performance of the inductor, such as the reduced inductance, increased resistance and capacitance. These effects are thoroughly investigated using the developed equivalent circuit model which is also utilized to design the pad-stacked inductor. The compact and broadband $\pi $ -diode with the reduced eddy current is then designed by the proper selection of the inductor layer and a patterned I/O pad. Measurement of the designed $\pi $ -diode with the pad-stacked inductor exhibits broadband impedance match and insertion loss, that is, a return loss better than 10 dB up to 26.5 GHz and a 3-dB bandwidth as large as 22.9 GHz. The chip area of the $\pi $ -diode remains the same as that of the I/O pad thanks to the proposed pad-stacked inductor. Therefore, the designed $\pi $ -diode can be applied for the compact high-speed I/O interface circuits.

Cryptocurrencies have recently attracted considerable attention, resulting in research mainly on deep learning-based price prediction models to maximize profit. Two research approaches have been adopted. Studies adopting the first approach directly predict the future cryptocurrency price. Long short-term memory (LSTM) and gated recurrent unit (GRU), which show high performance in time-series data, are mainly used for this approach. Further, studies adopting the second approach recommend actions to investors to maximize profits, such as “Sell”, “Buy”, and “Wait.” In this approach, classification models are used and results are derived based on probabilities. However, these action recommendation models do not consider the quality of the result. For example, it is risky to accept the result when the probability that the result of the action recommendation model for two classes is the correct answer is approximately 51%. To solve this problem, we recommend a method for adjusting the result of the action recommendation model based on Twitter sentiment analysis. The experimental results show that the proposed adjustment method improves the performance by approximately 3% compared to the conventional methods and are statistically validated.

To tackle a Doppler sensitivity problem of orthogonal frequency division multiplexing (OFDM), orthogonal time frequency space (OTFS) has been investigated, where information is carried over delay-Doppler domain. In this paper, to improve communication reliability in doubly dispersive channel, an auto-encoder (AE)-based OTFS modulation and detection scheme is developed, where the transmit OTFS waveform and its associated detection scheme at the receiver are jointly optimized in a deep learning framework. However, the conventional AE architecture which takes one-hot encoded input vector is hard to be reused in OTFS due to its enormous input dimensionality that increases exponentially on the number of grid points in delay-Doppler domain. To overcome it, we divide the delay-Doppler grid into multiple subblocks and associate the one-hot encoded vector with each subblock. Then, by concatenating them, one multi-hot vector is formed and exploited as the input vector for the proposed AE-based OTFS modulation and detection. We also develop a meta-learning scheme to effectively train the AE-based OTFS transceiver for newly updated channel profile.

In recent years, artificial intelligence (AI) has been applied in various fields, with rapid expansion of the scope of AI-human interactions. However, most AI technologies continue to exhibit black-box characteristics, i.e., their decisions and actions are not explainable, degrading user experience (UX). Recently, research on explainable AI (XAI), focusing on both AI performance and UX, has garnered significant attention. However, development of generalizable UX evaluation tools for AI and improvement of AI UX have not been investigated adequately. In this study, a UX evaluation tool is developed for AI based on a systematic literature review and verified using exploratory and confirmatory factor analyses. Subsequently, based on identified AI UX factors, UX of an AI defense system is upgraded. Based on user evaluation, there were significant improvements in eight factors (satisfaction, safety, controllability, trust, causality, efficiency, accountability, and explainability) with the exception of fairness. The proposed evaluation tool is expected to serve as a cornerstone for future evaluation of AI UX advancement.

Constructing a high dynamic range (HDR) image from multi-exposure low dynamic range (LDR) images is challenging mainly due to two major problems. One is a large misalignment between the LDR images taken at different moments due to camera and object motions. The other is missing content in over- or under-exposed areas, which can be very large in the LDR images with long or short exposure times. These problems lead to ghosting artifacts or saturated regions in the reconstructed HDR image. In this paper, we propose a method using convolutional neural networks (CNNs) to address these problems. Specifically, each LDR image is fed to a two-branch CNN that extracts two kinds of features separately, named content features and global features. Then, another CNN aligns and merges the features to create the HDR output. The content features are used to create spatial attention masks, which are used to align the features in the time domain with exposure-invariant attributes. Further, Fourier coefficient vectors of global features are utilized to modulate intermediate features in the frequency domain during the reconstruction process. Experimental results show that our method achieves state-of-the-art HDR reconstruction performances on several benchmarks. Our code is available at https://github.com/KeuntekLee/FDM-HDR.

In the autonomous car, perception with point cloud semantic segmentation helps obtain a wealth of information about the surrounding road environment. Despite the massive progress of recent researches, the existing machine learning networks are still insufficient for online applications of autonomous driving due to too subdivided classes, the lack of training data, and their heavy computing load. To solve these problems, we propose a fast and lite point cloud semantic segmentation network for autonomous driving, which utilizes LiDAR synthetic data to improve the performance by transfer learning. First, we modify the labeling classes and generate the LiDAR synthetic data-set for additional training to alleviate the lack of training data of the realistic data-set. Then, to lower the computing load, we adopt a projection-based method and apply a lightweight segmentation network to projected LiDAR images, which has drastically reduced computing. Finally, we verified and evaluated the proposed network in this paper through experiments. Experimental results show that the proposed network can perform the three-dimensional point cloud semantic segmentation in an online way, in which the inference speed overwhelms the existing algorithms.

Recently, the interest in the lower-limb exoskeleton robot is increasing within rehabilitation practice. While research on advancing the gait rehabilitation exoskeleton robot (ER) is being actively conducted from a technical point of view, studies evaluating usability and user value of ERs are insufficient. In this study, an evaluation tool was developed and verified, and an evaluation of an ER-based rehabilitation training program was conducted using the evaluation tool. A total of 102 rehabilitation patients (51 using ER and 51 under the existing training program [ETP] without robots) and 38 therapists participated in this experiment. The ER used in this study was Exowalk (HR-02, HMH Co. Ltd, South Korea). As a result of the study, the usability of the ERs consists of trust and overall satisfaction. The patient group using the ER showed statistically higher trust and overall satisfaction scores than the patient group with the ETP. On the other hand, the user value of the ER includes self-direction and security. The self-direction score of the patients using the ER was significantly higher than that of the patients with the ETP, but the security score showed the opposite tendency. This study is unique in that measures were developed to evaluate the usability and user value of the gait rehabilitation ER, especially for the Exowalk, verified by various user groups, and statistically analyzed to be used as a reference for the design of usability and user value in other gait rehabilitation ERs.

The uninterrupted operation of systems with artificial intelligence (AI) ensures high productivity and accuracy of the tasks performed. The physiological state of AI operators indicates a relationship with an AI system failure event and can be measured through electrodermal activity. This study aims to model the stress levels of system operators based on system trustworthiness and physiological responses during a correct AI operation and its failure. Two groups of 18 and 19 people participated in the experiments using two different types of software with elements of AI. The first group of participants used English proofreading software, and the second group used drawing software as the AI tool. During the tasks, the electrodermal activities of the participants as a stress level indicator were measured. Based on the results obtained, the users’ stress was determined and classified using logistic regression models with an accuracy of approximately 70%. The insights obtained can serve AI product developers in increasing the level of user trust and managing the anxiety and stress levels of AI operators.

Recently, the deep learning algorithm has received considerable attention and is influencing different fields including human-computer interaction (HCI). The purpose of this study is to maximize accuracy by applying deep learning to the classification of body movements. An experiment was performed to collect acceleration information on the wrist while performing seven workouts: pull up, row-barbell, bench press, dips, squat, deadlift, and military press. Participants were asked to perform each workout for ten sets repeated ten times per set. Experimental results confirm that one-dimensional convolutional neural network was the best among different algorithms including support vector machine, multi-layer perceptron, long short-term memory, and other deep convolutional neural networks. The accuracy was extremely high, 96%. The results of this experiment are applicable not only to the classification of fitness activities but also to the classification of different motions in numerous sporting events.

To increase the flexibility of network architecture, a mobile X-haul network (MXN) structure has been considered, in which the wireless front-haul, mid-haul, and backhaul links are integrated into a wireless X-haul link. Accordingly, the mobile station can be served by flexible transmission reception point (f-TRP) or remote radio head through the access link and the associated data should be delivered from XDU to the f-TRP through the X-haul link. Accordingly, the system throughput is determined by the minimum value of the access link capacity and the X-haul link capacity. In this paper, when millimeter-wave (mmWave) is exploited in such MXN with in-band full duplex, the optimal beamforming strategy is proposed, in which the receive beamforming for the X-haul link and the transmit beamforming for the access link are jointly designed to maximize the system throughput. Furthermore, we investigate the rate loss due to the quantization error of the channel state information at f-TRP with limited feedback, in which its upper bound is analytically derived and exploited in the determination of the number of feedback bits for a given finite rate loss.