RESUMO
Visually impaired people seek social integration, yet their mobility is restricted. They need a personal navigation system that can provide privacy and increase their confidence for better life quality. In this paper, based on deep learning and neural architecture search (NAS), we propose an intelligent navigation assistance system for visually impaired people. The deep learning model has achieved significant success through well-designed architecture. Subsequently, NAS has proved to be a promising technique for automatically searching for the optimal architecture and reducing human efforts for architecture design. However, this new technique requires extensive computation, limiting its wide use. Due to its high computation requirement, NAS has been less investigated for computer vision tasks, especially object detection. Therefore, we propose a fast NAS to search for an object detection framework by considering efficiency. The NAS will be used to explore the feature pyramid network and the prediction stage for an anchor-free object detection model. The proposed NAS is based on a tailored reinforcement learning technique. The searched model was evaluated on a combination of the Coco dataset and the Indoor Object Detection and Recognition (IODR) dataset. The resulting model outperformed the original model by 2.6% in average precision (AP) with acceptable computation complexity. The achieved results proved the efficiency of the proposed NAS for custom object detection.
Assuntos
Aprendizado Profundo , Tecnologia Assistiva , Auxiliares Sensoriais , Pessoas com Deficiência Visual , HumanosRESUMO
Facilitating the navigation of visually impaired people in indoor environments requires detecting indicating signs and informing them. In this paper, we proposed an indoor sign detection based on a lightweight anchor-free object detection model called FAM-centerNet. The baseline model of this work is the centerNet, which is an anchor-free object detection model with high performance and low computation complexity. A Foreground Attention Module (FAM) was introduced to extract target objects in real scenes with complex backgrounds. This module segments the foreground to extract relevant features of the target object using midground proposal and boxes-induced segmentation. In addition, the foreground module provides scale information to improve the regression performance. Extensive experiments on two datasets prove the efficiency of the proposed model for detecting general objects and custom indoor signs. The Pascal VOC dataset was used to test the performance of the proposed model for detecting general objects, and a custom dataset was used for evaluating the performance in detecting indoor signs. The reported results have proved the efficiency of the proposed FAM in enhancing the performance of the baseline model.
Assuntos
Tecnologia Assistiva , Pessoas com Deficiência Visual , Humanos , Algoritmos , AtençãoRESUMO
Biomarkers including fasting blood sugar, heart rate, electrocardiogram (ECG), blood pressure, etc. are essential in the heart disease (HD) diagnosing. Using wearable sensors, these measures are collected and applied as inputs to a deep learning (DL) model for HD diagnosis. However, it is observed that model accuracy weakens when the data gathered are scarce or imbalanced. Therefore, this work proposes two DL-based frameworks, GAN-1D-CNN, and GAN-Bi-LSTM. These frameworks contain: (1) a generative adversarial network (GAN) and (2) a one-dimensional convolutional neural network (1D-CNN) or bi-directional long short-term memory (Bi-LSTM). The GAN model is utilized to augment the small and imbalanced dataset, which is the Cleveland dataset. The 1D-CNN and Bi-LSTM models are then trained using the enlarged dataset to diagnose HD. Unlike previous works, the proposed frameworks increase the dataset first to avoid the prediction bias caused by the limited data. The GAN-1D-CNN achieved 99.1% accuracy, specificity, sensitivity, F1-score, and 100% area under the curve (AUC). Similarly, the GAN-Bi-LSTM obtained 99.3% accuracy, 99.2% specificity, 99.3% sensitivity, 99.2% F1-score, and 100% AUC. Furthermore, time complexity of proposed frameworks is investigated with and without principal component analysis (PCA). The PCA method reduced prediction times for 61 samples using GAN-1D-CNN and GAN-Bi-LSTM to 68.8 and 74.8 ms, respectively. These results show that it is reliable to use our frameworks for augmenting limited data and predicting heart disease.
RESUMO
The emerging novel variants and re-merging old variants of SARS-CoV-2 make it critical to study the transmission probability in mixed-mode ventilated office environments. Artificial neural network (ANN) and curve fitting (CF) models were created to forecast the R-Event. The R-Event is defined as the anticipated number of new infections that develop in particular events occurring over the course of time in any defined space. In the spring and summer of 2022, real-time data for an office environment were collected in India in a mixed-mode ventilated office space in a composite climate. The performances of the proposed CF and ANN models were compared with respect to traditional statistical indicators, such as the correlation coefficient, RMSE, MAE, MAPE, NS index, and a20-index, in order to determine the merit of the two approaches. Thirteen input features, namely the indoor temperature (TIn), indoor relative humidity (RHIn), area of opening (AO), number of occupants (O), area per person (AP), volume per person (VP), CO2 concentration (CO2), air quality index (AQI), outer wind speed (WS), outdoor temperature (TOut), outdoor humidity (RHOut), fan air speed (FS), and air conditioning (AC), were selected to forecast the R-Event as the target. The main objective was to determine the relationship between the CO2 level and R-Event, ultimately producing a model for forecasting infections in office building environments. The correlation coefficients for the CF and ANN models in this case study were 0.7439 and 0.9999, respectively. This demonstrates that the ANN model is more accurate in R-Event prediction than the curve fitting model. The results show that the proposed ANN model is reliable and significantly accurate in forecasting the R-Event values for mixed-mode ventilated offices.
Assuntos
Poluição do Ar em Ambientes Fechados , COVID-19 , Humanos , SARS-CoV-2 , Dióxido de Carbono , COVID-19/epidemiologia , Clima , Redes Neurais de Computação , Poluição do Ar em Ambientes Fechados/análise , VentilaçãoRESUMO
It is critical to successfully identify, mitigate, and fight against Android malware assaults, since Android malware has long been a significant threat to the security of Android applications. Identifying and categorizing dangerous applications into categories that are similar to one another are especially important in the development of a safe Android app ecosystem. The categorization of malware families may be used to improve the efficiency of the malware detection process as well as to systematically identify malicious trends. In this study, we proposed a modified ResNeXt model by embedding a new regularization technique to improve the classification task. In addition, we present a comprehensive evaluation of the Android malware classification and detection using our modified ResNeXt. The nonintuitive malware's features are converted into fingerprint images in order to extract the rich information from the input data. In addition, we applied fine-tuned deep learning (DL) based on the convolutional neural network (CNN) on the visualized malware samples to automatically obtain the discriminatory features that separate normal from malicious data. Using DL techniques not only avoids the domain expert costs but also eliminates the frequent need for the feature engineering methods. Furthermore, we evaluated the effectiveness of the modified ResNeXt model in the classification process by testing a total of fifteen different combinations of the Android malware image sections on the Drebin dataset. In this study, we only use grayscale malware images from a modified ResNeXt to analyze the malware samples. The experimental results show that the modified ResNeXt successfully achieved an accuracy of 98.25% using Android certificates only. Furthermore, we undertook extensive trials on the dataset in order to confirm the efficacy of our methodology, and we compared our approach with several existing methods. Finally, this article reveals the evaluation of different models and a much more precise option for malware identification.
Assuntos
Ecossistema , Redes Neurais de Computação , HumanosRESUMO
Many hardware and software advancements have been made to improve image quality in smartphones, but unsuitable lighting conditions are still a significant impediment to image quality. To counter this problem, we present an image enhancement pipeline comprising synthetic multi-image exposure fusion and contrast enhancement robust to different lighting conditions. In this paper, we propose a novel technique of generating synthetic multi-exposure images by applying gamma correction to an input image using different values according to its luminosity for generating multiple intermediate images, which are then transformed into a final synthetic image by applying contrast enhancement. We observed that our proposed contrast enhancement technique focuses on specific regions of an image resulting in varying exposure, colors, and details for generating synthetic images. Visual and statistical analysis shows that our method performs better in various lighting scenarios and achieves better statistical naturalness and discrete entropy scores than state-of-the-art methods.