Recent regulations to block the widespread transmission of COVID-19 disease among people impose the use of facial masks indoor and outdoor. Such restriction becomes critical in all those scenarios where access controls take benefit from biometric recognition systems. The occlusions due to the presence of a facial mask make a significant portion of human faces unavailable for feature extraction and analysis. This work explores the contribution of the solely periocular region of the face to achieve a robust recognition approach suitable for mobile devices. Rather than working on a static analysis of the facial features, like largely done by work on periocular recognition in the literature, the proposed study focuses the attention on the analysis of face dynamics so that the spatio-temporal features make the recogniser frame-independent and tolerant to user movements during the acquisition. To obtain a lightweight processing, which is compliant with limited computing power of mobile devices, the spatio-temporal representation of the periocular region has analysed and classified through Machine Learning approaches. The experimental discussion has been performed on a new dataset, Mobile Masked Face REcognition Database, specifically designed to analyse the periocular region dynamics in presence of facial masks. For a wider comparative analysis, a publicly available dataset called XM2VTS has been considered as well as Deep Learning solutions have been experimented to discuss the challenging aspects of the recognition problem. Moreover, a summary of the state-of-the-art on periocular recognition driven by COVID pandemic has been presented, showing how the research efforts in this field focused on recognition of still images. Experimental results show promising levels of performance as well as limitations of the proposed approach, creating the premises for future directions.

https://ieeexplore.ieee.org/ielx7/6287639/6514899/09874825.pdf

M2FRED: Mobile Masked Face REcognition Through Periocular Dynamics Analysis

Adversarial attacks rely on the instability phenomenon appearing in general for all inverse problems, e.g., image classification and reconstruction, independently of the computational scheme or method used to solve the problem. We mathematically prove and empirically show that machine learning denoisers (MLD) are not excluded. That is to prove the existence of adversarial attacks given by noise patterns making the MLD run into instability, i.e., the MLD increases the noise instead of decreasing it. We further demonstrate that adversarial retraining or classic filtering do not provide an exit strategy for this dilemma. Instead, we show that adversarial attacks can be inferred by polynomial regression. Removing the underlying inferred polynomial distribution from the total noise distribution delivers an efficient technique yielding robust MLDs that make consistent computer vision tasks such as image segmentation or classification more reliable.

Adversarial attacks for machine learning denoisers and how to resist them

Continuous, real-time monitoring of occupational health and safety in high-risk workplaces such as construction sites can substantially improve the safety of workers. However, introducing such systems in practice is associated with a number of challenges, such as scaling up the solution while keeping its cost low. In this context, this work investigates the use of an off-the-shelf, low-cost smartwatch to detect health issues based on heart rate monitoring in a privacy-preserving manner. To improve the smartwatch’s low measurement quality, a novel, frugal machine learning method is proposed that corrects measurement errors, along with a new dataset for this task. This method’s integration with the smartwatch and the remaining parts of the health and safety monitoring system (built on the ASSIST-IoT reference architecture) are presented. This method was evaluated in a laboratory environment in terms of its accuracy, computational requirements, and frugality. With an experimentally established mean absolute error of 8.19 BPM, only 880 bytes of required memory, and a negligible impact on the performance of the device, this method meets all relevant requirements and is expected to be field-tested in the coming months. To support reproducibility and to encourage alternative approaches, the dataset, the trained model, and its implementation on the smartwatch were published under free licenses.

Frugal Heart Rate Correction Method for Scalable Health and Safety Monitoring in Construction Sites

Photoplethysmography (PPG) is a simple and cost-efficient technique that effectively measures cardiovascular response by detecting blood volume changes in a noninvasive manner. A practical challenge in the use of PPGs in real-world applications is noise reduction. PPG signals are likely to be compromised by various types of noise, such as scattering or motion artifacts, and removing such compounding noises using a monotonous method is not easy. To this end, this paper proposes a neural PPG denoiser that can robustly remove multiple types of noise from a PPG signal. By casting the noise reduction problem into a signal restoration approach, we aim to achieve a solid performance in the reduction of different noise types using a single neural denoiser built upon transformer-based deep generative models. Using this proposed method, we conducted the experiments on the noise reduction of a PPG signal synthetically contaminated with five types of noise. Following this, we performed a comparative study using six different noise reduction algorithms, each of which is known to be the best model for each noise. Evaluation results of the peak signal-to-noise ratio (PSNR) show that the neural PPG denoiser is superior in three out of five noise types to the performance of conventional noise reduction algorithms. The salt-and-pepper noise type showed the best performance, with the PSNR of the neural PPG denoiser being 36.6080, and the PSNRs of the other methods were 19.8160 and 32.8234. The Poisson noise type performed the worst, showing a PSNR of 33.0090; the PSNRs of other methods were 35.1822 and 33.4795, respectively. Thereafter, an experiment to recover a signal synthesized with two or more of the five noise types was conducted. When the number of mixed noises was two, three, four, and five, the PSNRs were 29.2759, 27.8759, 26.5608, and 25.9402, respectively. Finally, an experiment to recover motion artifacts was also conducted. The synthesized motion artifact signal was created by synthesizing only a certain ratio of the total signal length. As a result of the motion artifact signal restoration, the PSNRs were 25.2872, 22.8240, 21.2901, and 19.9577 at 30%, 50%, 70%, and 90% motion artifact ratios, respectively. In the three experiments conducted, the neural PPG denoiser showed that various types of noise were effectively removed. This proposal contributes to the universal denoising of continuous PPG signals and can be further expanded to denoise continuous signals in the general domain.

/pdf/preeminently-robust-neural-ppg-denoiser-3ltr0k5r.pdf

Preeminently Robust Neural PPG Denoiser

Due to COVID-19 pandemic, wearing a mask rapidly becomes a new social norm that people in the society should comply Although it is altruistic behavior preventing all people from serious infections, it brings about hassles for individuals For example, when a face is partly covered with a mask, identification using a face recognition is going to be malfunctioning To this end, we propose a novel computational framework that enables the personal authentication with a partial face image, a face covered with a mask For the experiments, we constructed the datasets of facial images containing the periocular regions only, extracted from full facial images covered with the mask Given the datasets, we trained our framework, a variant of a Siamese network, with various configuration of hyper-parameters As a result, RMSprop optimizer with the learning rate 1 \(\times \) 10–5 trained from periocular datasets showed the highest accuracy for the personal authentication Next, we conducted a comparative experiment with our proposal and the model trained with datasets containing the full facial regions When testing with the periocular region images, our proposal is superior in the authentication accuracy to that of the model trained with the full facial regions This result raises the optimistic expectation that in the era of COVID-19, facial images covered with mask can still be used for the authentication using face recognition at a nearly same level of accuracy This means that people can use the face recognition applications without taking off the mask, which provides the safe circumstances against the infections

Authentication of Facial Images with Masks Using Periocular Biometrics

Improving Gaze Estimation Performance Using Ensemble Loss Function

In services such as metaverse, which should provide a constant quality of service (QoS) regardless of the user’s physical location, the end-to-end (E2E) latency must be fairly distributed over any flow in the network. To this end, we propose a reinforcement learning (RL)-based scheduler for minimizing the maximum network E2E latency. The RL model used the double deep Q-network (DDQN) with the prioritized experience replay (PER). In order to see the performance change according to the type of RL agent, we implemented a single-agent environment where the controller is an agent and a multi-agent environment where each node is an agent. Since the agents were unable to identify E2E latencies in the multi-agent environment, the state and reward were formulated using the estimated E2E latencies. To precisely evaluate the RL-based scheduler, we set out benchmark algorithms to compare with which a network-arrival-time-based heuristic algorithm (NAT-HA) and a maximum-estimated-delay-based heuristic algorithm (MED-HA). The RL-based scheduler, first-in-first-out (FIFO), round-robin (RR), NAT-HA, and MED-HA were compared through large-scale simulations on four network topologies. The simulation results in fixed-packet generation scenarios showed that our proposal, the RL-based scheduler, achieved the minimization of maximum E2E latency in all the topologies. In other scenarios with random flow generation, the RL-based scheduler and MED-HA showed the lowest maximum E2E latency for all topologies. Depending on the topology, the maximum E2E latency of NAT-HA was equal to or larger than that of the RL-based scheduler. In terms of fairness, the RL-based scheduler showed a higher level of fairness than that of FIFO and RR. NAT-HA had similar or lower fairness than the RL-based scheduler depending on the topology, and MED-HA had the same level of fairness as the RL-based scheduler.

/pdf/improving-end-to-end-latency-fairness-using-a-reinforcement-laodwhrk.pdf

Jeeghang Lee

Papers

Preeminently Robust Neural PPG Denoiser

Authentication of Facial Images with Masks Using Periocular Biometrics

Improving Gaze Estimation Performance Using Ensemble Loss Function

Improving End-To-End Latency Fairness Using a Reinforcement-Learning-Based Network Scheduler