Reinforced Cascading Convolutional Neural Networks and Vision Transformer for Lung Disease Diagnosis

Lung diseases are among the most deadly infectious diseases worldwide. Covid-19 infection is a current disease that falls within this category and has impacted public health in countries across the globe. Accordingly, this study focuses on building a lung disease identification system using a state-of-the-art deep cascade learning classification model, EfficientNet-Vision Transformer. The proposed Real ESRGAN is utilized to enhance the input of EfficientNet, while image Relative Position Encoding (iRPE) is added to improve the attention of the transformer network. Moreover, weight balancing is applied to stabilize the performance of the proposed system. When trained on the X-Ray dataset, our model achieved 93.757% accuracy on five classes of lung disease: Normal, Covid-19, Viral Pneumonia, Bacterial Pneumonia, and Tuberculosis. © 2022 IEEE.

An Improved Face Mask-aware Recognition System Based on Deep Learning

Face mask detection and recognition have been incorporated into many applications in daily life, especially during the current COVID-19 pandemic. To mitigate the spread of coronavirus, wearing face masks has become commonplace. However, traditional face detection and recognition systems utilize main facial features such as the mouth, nose, and eyes to determine a person’s identity. Masks make facial detection and recognition tasks more challenging since certain parts of the face are concealed. Yet, how to improve the performance of existing systems with a face mask overlaid on the original face input images remains an open area of inquiry. In this study, we propose an improved face mask-aware recognition system named ‘MAR’ based on deep learning, which can tackle challenges in face mask detection and recognition. MAR consists of five main modules to handle various kinds of input images. We re-train the CenterNet model with our augmented face mask inputs to perform face mask detection and propose four variations on face mask recognition models based on the pre-trained ArcFace to handle facial recognition. Finally, we demonstrate the effectiveness of our proposed models on the VGGFACE2 dataset and achieve a high accuracy score on both detection and recognition tasks. © 2022, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.