Adaptive Spatial Transformation Networks for Periocular Recognition.

Periocular recognition has emerged as a particularly valuable biometric identification method in challenging scenarios, such as partially occluded faces due to COVID-19 protective masks masks, in which face recognition might not be applicable. This work presents a periocular recognition framework based on deep learning, which automatically localises and analyses the most important areas in the periocular region. The main idea is to derive several parallel local branches from a neural network architecture, which in a semi-supervised manner learn the most discriminative areas in the feature map and solve the identification problem solely upon the corresponding cues. Here, each local branch learns a transformation matrix that allows for basic geometrical transformations (cropping and scaling), which is used to select a region of interest in the feature map, further analysed by a set of shared convolutional layers. Finally, the information extracted by the local branches and the main global branch are fused together for recognition. The experiments carried out on the challenging UBIRIS-v2 benchmark show that by integrating the proposed framework with various ResNet architectures, we consistently obtain an improvement in mAP of more than 4% over the "vanilla" architecture. In addition, extensive ablation studies were performed to better understand the behavior of the network and how the spatial transformation and the local branches influence the overall performance of the model. The proposed method can be easily adapted to other computer vision problems, which is also regarded as one of its strengths.

A feasible route for the design and manufacture of customised respiratory protection through digital facial capture.

The World Health Organisation has called for a 40% increase in personal protective equipment manufacturing worldwide, recognising that frontline workers need effective protection during the COVID-19 pandemic. Current devices suffer from high fit-failure rates leaving significant proportions of users exposed to risk of viral infection. Driven by non-contact, portable, and widely available 3D scanning technologies, a workflow is presented whereby a user's face is rapidly categorised using relevant facial parameters. Device design is then directed down either a semi-customised or fully-customised route. Semi-customised designs use the extracted eye-to-chin distance to categorise users in to pre-determined size brackets established via a cohort of 200 participants encompassing 87.5% of the cohort. The user's nasal profile is approximated to a Gaussian curve to further refine the selection in to one of three subsets. Flexible silicone provides the facial interface accommodating minor mismatches between true nasal profile and the approximation, maintaining a good seal in this challenging region. Critically, users with outlying facial parameters are flagged for the fully-customised route whereby the silicone interface is mapped to 3D scan data. These two approaches allow for large scale manufacture of a limited number of design variations, currently nine through the semi-customised approach, whilst ensuring effective device fit. Furthermore, labour-intensive fully-customised designs are targeted as those users who will most greatly benefit. By encompassing both approaches, the presented workflow balances manufacturing scale-up feasibility with the diverse range of users to provide well-fitting devices as widely as possible. Novel flow visualisation on a model face is presented alongside qualitative fit-testing of prototype devices to support the workflow methodology.

AAN-Face: Attention Augmented Networks for Face Recognition.

Convolutional neural networks are capable of extracting powerful representations for face recognition. However, they tend to suffer from poor generalization due to imbalanced data distributions where a small number of classes are over-represented (e.g. frontal or non-occluded faces) and some of the remaining rarely appear (e.g. profile or heavily occluded faces). This is the reason why the performance is dramatically degraded in minority classes. For example, this issue is serious for recognizing masked faces in the scenario of ongoing pandemic of the COVID-19. In this work, we propose an Attention Augmented Network, called AAN-Face, to handle this issue. First, an attention erasing (AE) scheme is proposed to randomly erase units in attention maps. This well prepares models towards occlusions or pose variations. Second, an attention center loss (ACL) is proposed to learn a center for each attention map, so that the same attention map focuses on the same facial part. Consequently, discriminative facial regions are emphasized, while useless or noisy ones are suppressed. Third, the AE and the ACL are incorporated to form the AAN-Face. Since the discriminative parts are randomly removed by the AE, the ACL is encouraged to learn different attention centers, leading to the localization of diverse and complementary facial parts. Comprehensive experiments on various test datasets, especially on masked faces, demonstrate that our AAN-Face models outperform the state-of-the-art methods, showing the importance and effectiveness.

A numerical analysis of skin-PPE interaction to prevent facial tissue injury.

The use of close-fitting PPE is essential to prevent exposure to dispersed airborne matter, including the COVID-19 virus. The current pandemic has increased pressure on healthcare systems around the world, leading to medical professionals using high-grade PPE for prolonged durations, resulting in device-induced skin injuries. This study focuses on computationally improving the interaction between skin and PPE to reduce the likelihood of discomfort and tissue damage. A finite element model is developed to simulate the movement of PPE against the face during day-to-day tasks. Due to limited available data on skin characteristics and how these vary interpersonally between sexes, races and ages, the main objective of this study was to establish the effects and trends that mask modifications have on the resulting subsurface strain energy density distribution in the skin. These modifications include the material, geometric and interfacial properties. Overall, the results show that skin injury can be reduced by using softer mask materials, whilst friction against the skin should be minimised, e.g. through use of micro-textures, humidity control and topical creams. Furthermore, the contact area between the mask and skin should be maximised, whilst the use of soft materials with incompressible behaviour (e.g. many elastomers) should be avoided.

One size fits all?: A simulation framework for face-mask fit on population-based faces.

The use of face masks by the general population during viral outbreaks such as the COVID-19 pandemic, although at times controversial, has been effective in slowing down the spread of the virus. The extent to which face masks mitigate the transmission is highly dependent on how well the mask fits each individual. The fit of simple cloth masks on the face, as well as the resulting perimeter leakage and face mask efficacy, are expected to be highly dependent on the type of mask and facial topology. However, this effect has, to date, not been adequately examined and quantified. Here, we propose a framework to study the efficacy of different mask designs based on a quasi-static mechanical model of the deployment of face masks onto a wide range of faces. To illustrate the capabilities of the proposed framework, we explore a simple rectangular cloth mask on a large virtual population of subjects generated from a 3D morphable face model. The effect of weight, age, gender, and height on the mask fit is studied. The Centers for Disease Control and Prevention (CDC) recommended homemade cloth mask design was used as a basis for comparison and was found not to be the most effective design for all subjects. We highlight the importance of designing masks accounting for the widely varying population of faces. Metrics based on aerodynamic principles were used to determine that thin, feminine, and young faces were shown to benefit from mask sizes smaller than that recommended by the CDC. Besides mask size, side-edge tuck-in, or pleating, of the masks as a design parameter was also studied and found to have the potential to cause a larger localized gap opening.

A novel fully automatic design approach of a 3D printed face specific mask: Proof of concept.

The use of high quality facemasks is indispensable in the light of the current COVID pandemic. This study proposes a fully automatic technique to design a face specific mask. Through the use of stereophotogrammetry, computer-assisted design and three-dimensional (3D) printing, we describe a protocol for manufacturing facemasks perfectly adapted to the individual face characteristics. The face specific mask was compared to a universal design of facemask and different filter container's designs were merged with the mask body. Subjective assessment of the face specific mask demonstrated tight closure at the nose, mouth and chin area, and permits the normal wearing of glasses. A screw-drive locking system is advised for easy assembly of the filter components. Automation of the process enables high volume production but still allows sufficient designer interaction to answer specific requirements. The suggested protocol can be used to provide more comfortable, effective and sustainable solution compared to a single use, standardized mask. Subsequent research on printing materials, sterilization technique and compliance with international regulations will facilitate the introduction of the face specific mask in clinical practice as well as for general use.