Wearable assistant devices play an important role in daily life for people with disabilities. Those who are hearing impaired may face dangers while walking or driving on the road. The major danger is their inability to hear warning sounds from cars or ambulances. Thus, the goal of this study is to develop a wearable assistant device for the hearing impaired to recognize emergency vehicle sirens on the road using edge computing. An EfficientNet-based fuzzy rank-based ensemble model was proposed to classify seven audio sounds, including human vocalizations and emergency vehicle sirens. This model was embedded in an Arduino Nano 33 BLE Sense development board. The audio files were respectively obtained from the CREMA-D dataset and Large Scale Audio dataset of emergency vehicle sirens on the road, with a total number of 8756 files. The seven audio sounds included neutral vocalization, anger vocalization, fear vocalization, happy vocalization, car horn sound, siren sound, and ambulance siren sound. The audio signal was converted into a spectrogram by the short-time Fourier transform as the feature. When one of the car horns, sirens, or ambulance sirens was detected, the wearable assistant device presented alarms through vibration and messages on the OLED panel. The performances of the EfficientNet-based fuzzy rank-based ensemble model in offline computing achieved an accuracy of 97.1%, precision of 97.79%, sensitivity of 96.8%, and specificity of 97.04%. In edge computing, the results were an accuracy of 95.2%, precision of 93.2%, sensitivity of 95.3%, and specificity of 95.1%. Thus, the proposed wearable assistant device has the potential benefit of helping the hearing impaired avoid traffic accidents.

Document classification is one of the most critical steps in the document analysis pipeline. There are two types of approaches for document classification, known as image-based and multimodal approaches. The image-based document classification approaches are solely based on the inherent visual cues of the document images. In contrast, the multimodal approach co-learns the visual and textual features, and it has proved to be more effective. Nonetheless, these approaches require a huge amount of data. This paper presents a novel approach for document classification that works with a small amount of data and outperforms other approaches. The proposed approach incorporates a hierarchical attention network(HAN) for the textual stream and the EfficientNet-B0 for the image stream. The hierarchical attention network in the textual stream uses the dynamic word embedding through fine-tuned BERT. HAN incorporates both the word level and sentence level features. While the earlier approaches rely on training on a large corpus (RVL-CDIP), we show that our approach works with a small amount of data (Tobacco-3482). To this end, we trained the neural network at Tobacco-3428 from scratch. Thereby, we outperform state-of-the-art by obtaining an accuracy of 90.3%. This results in a relative error reduction rate of 7.9%.

A machine learning method for classifying Lung UltraSound is here proposed to pro- vide a point of care tool for supporting a safe, fast and accurate diagnosis, that can also be useful during a pandemic like as SARS-CoV-2. Given the advantages (e.g. safety, rapidity, portability, cost-effectiveness) provided by the ultrasound technology over other methods (e.g. X-ray, computer tomography, magnetic resonance imaging), our method was validated on the largest LUS public dataset. Focusing on both accuracy and efficiency, our solution is based on an efficient adaptive ensembling of two EfficientNet-b0 models reaching 100% of accuracy, which, to our knowledge, outperforms the previous state-of-the-art. The complexity of this solution keeps the number of parameters in the same order as an EfficientNet-b0 by adopting specific design choices that are adaptive ensembling with a combination layer, ensembling performed on the deep features, minimal ensemble only two weak models. Moreover, a visual analysis of the saliency maps on sample images of all the classes of the dataset reveals where the focus is on an inaccurate weak model versus an accurate model.

Automated extraction of buildings from earth observation (EO) data is important for various applications, including updating of maps, risk assessment, urban planning, policy making. Combining data from different sensors such as high-resolution multispectral (HRI) and light detection and ranging (LiDAR) has shown great potential in building extraction. Deep learning (DL) is increasingly used in multimodal data fusion and urban object extraction. However, DL-based multimodal fusion networks may underperform due to insufficient learning of “joint features” from multiple sources and oversimplified approaches to fusing multimodal features. Recently, an hybrid attention-aware fusion network (HAFNet) has been proposed for building extraction from a dataset including co-located Very-High-Resolution (VHR) optical images and Light Detection And Ranging (LiDAR) joint data. The system reported good performances thanks to the adaptivity of the attention mechanism to the features of the information content of the three streams but suffered from model overparametrization, which inevitably leads to long training times and heavy computational load. In this paper the authors propose a restructuring of the scheme, which involved replacing VGG-16-like encoders with the recently proposed EfficientNet, whose advantages counteract exactly the issues found with the HAFNet scheme. The novel configuration was tested on multiple benchmark datasets, reporting great improvements in terms of processing times, and also in terms of accuracy. The new scheme, called HAFNetE (HAFNet with EfficientNet integration), appears indeed capable of achieving good results with less parameters, translating into better computational efficiency. Based on these findings, we can conclude that, given the current advancements in single-thread schemes, the classical multi-thread HAFNet scheme could be effectively transformed by the HAFNetE scheme by replacing VGG-16 with EfficientNet blocks on each single thread. The remarkable reduction achieved in computational requirements moves the system one step closer to on-board implementation in a possible, future “urban mapping” satellite constellation.