Road surface monitoring in winter conditions is of great importance to ensure the safety of road users. Estimation of snow coverage on roads can be included in intelligent transportation systems to alert drivers or improve snow removal processes. Several models have been proposed for estimating snow coverage using surveillance cameras, but these models have focused on predicting few snow levels, which limits their usefulness in practice. In this paper, we present a model that allows a more granular estimation of the percentage of road surface covered by snow by predicting snow coverage from 0% (no snow) to 100% (fully snow-covered) using increments of 10%. We propose an ensemble learning model combining a deep convolutional neural network (CNN) and a support-vector machine (SVM). The accuracy of our model is similar to the state-of-the-art accuracy despite the higher task complexity associated with the increased granularity of predictions.

Wildfires are a worldwide natural risk, which causes harmful effects to human safety and leads to ecological and economical damage. Various fire detection systems have been proposed in order to detect fire and reduce its effects. However, they are still limited in detecting small fire areas and determining the precise fire’s shape. In order to overcome these limitations, we present, in this paper, a novel method based on deep learning, called ‘Deep-RegSeg’, to segment fire pixels and detect fire areas in complex non-structured environments. Deep-RegSeg is evaluated with varying backbone and loss function. The obtained results showed a high performance and outperformed some recent state-of-the-art techniques. The results also proved that Deep-RegSeg is efficient in segmenting wildfire pixels and detecting the precise fire’s shape, especially small fire areas under various conditions of weather, presence of smoke, and environment brightness.

Melanoma remains the most dangerous form of skin cancer which has a high mortality rate. When detect early, melanoma can be easily cured and millions of lives might be saved. The use of automatic detection models in clinical decision support can increase the ability to address this issue and improve survival rates. In this work, we proposed an automated pipeline for melanoma detection, which combines the predictions of deep convolutional neural network models through ensemble learning techniques. Furthermore, our automated pipeline includes various strategies such as image augmentation, upsampling, image cropping, digital hair removal and class weighting. Our pipeline was trained and tested using the image data acquired from the Society for Imaging Informatics in Medicine and the International Skin Imaging Collaboration SIIM-ISIC 2020. Our proposed pipeline has demonstrated a high performance compared to the other state-of-the-art pipelines for melanoma disease prediction with an accuracy of 97.77% and an AUC of 98.47%.

Chest diseases are among the most common worldwide health problems; they are potentially life-threatening disorders which can affect organs such as lungs and heart. Radiologists typically use visual inspection to diagnose chest X-ray (CXR) diseases, which is a difficult task prone to errors. The signs of chest abnormalities appear as opacities around the affected organ, making it difficult to distinguish between diseases of superimposed organs. To this end, we propose a very first method for CXR organ disease detection using deep learning. We used an ensemble learning (EL) approach to increase the efficiency of the classification of CXR diseases by organs (lung and heart) using a consolidated dataset. This dataset contains 26,316 CXR images from VinDr-CXR and CheXpert datasets. The proposed ensemble of deep convolutional neural networks (DCNN) approach achieves excellent performance with an AUC of 0.9489 for multi-class classification, outperforming many state-of-the-art models.

Human body parts detection is an important field of research in computer vision. It can serve as an essential tool in surveillance systems and used to automatically detect and moderate non-appropriate online content such as nudity, child pornography, violence, etc. In this work, we introduce a novel two-step framework to define ten body parts using joints localization. A new architecture with EfficientNet as a backbone is proposed and compared to HRNet for the first step of pose estimation. The resulting joints are then used as an input to the second step, where a set of rules is applied to connect the appropriate joints and to define each body part. The developed algorithms were tested using MPII human pose benchmark. The proposed approach achieved a very interesting performance with a 90.13% Probability of Correct Keypoint (PCK) for the pose estimation and an average of 89.80% of mean Average Precision (mAP) for the body parts detection.

The Coronavirus Disease 2019 (COVID-19) pandemic continues to have a devastating effect on the global population’s health and well-being. Successful screening of infected patients is a crucial step in the battle against COVID-19, with radiology inspection using chest radiography being one of the most popular screening methods. Early studies discovered that patients with COVID-19 infection have anomalies in chest radiography images. In this study, we present our Deep Convolutional Neural Network (CNN) for an automatic detection of COVID-19 using computed tomography (CT). Multiple models are presented and fine-tuned to provide accurate detection of COVID-19 vs. normal vs. pneumonia. The proposed model gives an Area Under Curve (AUC) of 99.64%, an accuracy (ACC) of 96.37%, a specificity of 96.00% and a sensitivity of 97.00%. Moreover, an explainability algorithm has been developed and shows the high efficiency of identifying the pathological signs of COVID-19 in CT scans.

Forest fires are among the most dangerous type of natural disasters since they affect numerous aspects of life, such as natural ecosystems, economy, and human lives. Various vision-based fire detection methods have been proposed to segment fire pixels and detect fire at an early stage. The challenge here is to overcome the limitations of the majority of these methods mainly false detection of fire pixels. For such, we propose in this paper, three deep convolutional networks, U-Net, U2-Net, and EfficientSeg to segment forest fire pixels and detect fire areas. One of our main contributions is the variation of loss functions of all models. The three models show an excellent performance in terms of accuracy and F1-score, and proved their reliability to segment fire pixels and detect the precise shape of forest fire areas.

In recent years we have witnessed an increase in cyber threats and malicious software attacks on different platforms with important consequences to persons and businesses. It has become critical to find automated machine learning techniques to proactively defend against malware. Transformers, a category of attention-based deep learning techniques, have recently shown impressive results in solving different tasks mainly related to the field of Natural Language Processing (NLP). In this paper, we propose the use of a Transformers architecture to automatically detect malicious software. We propose MalBERT, a model based on BERT (Bidirectional Encoder Representations from Transformers) which performs a static analysis on the source code of Android applications using preprocessed features to characterize existing malware and classify it into different representative malware categories. The obtained results are promising and show the high performance obtained by Transformer-based models for malicious software detection.

The important growth of social media and online gaming sites in recent years have increased the challenge of online moderation to keep the internet safe and without toxic content. Today, machine learning techniques play an important role in detecting inappropriate content and help moderate online interaction. Text classification using Natural Language Processing (NLP) methods has been extensively studied using deep learning models and transformers which have shown impressive results. Despite this, specific classification tasks on limited datasets still need to be improved. In this paper, we propose an approach based on an Attention-Based Bidirectional LSTM model and a combination of custom features to enhance automatic misogyny identification (AMI) on social media. We present a multi-lingual study of the phenomena by carrying out different classification experiments. Our study focuses on selecting most important features to improve the model for misogyny detection. The proposed model outperforms many state-of-the-art approaches across multiple datasets.

Violence and abnormal behavior detection research have known an increase of interest in recent years, due mainly to a rise in crimes in large cities worldwide. In this work, we propose a deep learning architecture for violence detection, which combines both recurrent neural networks (RNNs) and 2-dimensional convolutional neural networks (2D CNN). In addition to video frames, we use optical flow computed using the captured sequences. CNN extracts spatial characteristics in each frame, while RNN extracts temporal characteristics. The use of optical flow allows to encode the movements in the scenes. The proposed approaches reach the same level as state-of-the-art techniques and sometimes surpass them. The techniques were validated on three databases achieving very interesting results.

Recently, computer vision‐based methods have started to replace conventional sensor‐based fire detection technologies. In general, visible band image sequences are used to automatically detect suspicious fire events in indoor or outdoor environments. There are several methods which aim to achieve automatic fire detection on visible band images, however, it is difficult to identify which method is the best performing as there is no fire image dataset which can be used to test the different methods. This study presents a benchmarking of state of the art wildland fire colour segmentation algorithms using a new fire dataset introduced for the first time. The dataset contains images of wildland fire in different contexts (fuel, background, luminosity, smoke etc.). All images of the dataset are characterised according to the principal colour of the fire, the luminosity, and the presence of smoke in the fire area. With this characterisation, it has been possible to determine on which kind of images each algorithm is efficient. Also a new probabilistic fire segmentation algorithm is introduced and compared to the other techniques. Benchmarking is performed in order to assess performances of 12 algorithms that can be used for the segmentation of wildland fire images.

This work present a framework for the evaluation of face recognition performances in and beyond the visible spectrum. A new multispectral face database was developed and includes visible and different infrared spectrums. Face extraction and normalization techniques are introduced and used to normalize face images from the Laval University and Equinox databases. This framework propose different face learning and recognition techniques. Also, Texture space transformations are implemented and can be used prior to face learning and recognition. Finally, multi-scale fusion algorithms are present in this framework and can be used for the evaluation of fusion schemes in multispectral face recognition.

This work introduces a new locally adaptive texture features for efficient multispectral face recognition. This new descriptor called Local Adaptive Ternary Pattern (LATP) is based on the Local Ternary Pattern (LTP). Unlike the previous techniques, this new descriptor determines the local pattern threshold automatically using local statistics. It shares with LTP the property of being less sensitive to noise, illumination change and facial expressions. These characteristics make it a good candidate for multispectral face recognition. Linear and non linear subspace learning and recognition techniques are introduced and used for performance evaluation of face recognition in the new texture space: PCA, LDA, Kernel-PCA (KPCA), Kernel-LDA (KDA), Linear Graph Embedding (LGE), Kernel-LGE (KLGE), Locality Preserving Projection (LPP) and Kernel-LPP (KLPP). The obtained results show an increase in recognition performance when texture features are used. LTP and LATP are the best performing techniques. The overall best performance is obtained in the short wave infrared spectrum (SWIR) using the new proposed technique combined with a non linear subspace learning technique.

In this work we present an effective framework for color-texture classification where statistical features are computed from a generalized isotropic co-occurrence matrix extracted from color bands and combined with image entropies. The proposed approach has been effectively tested in RGB, HSL and La*b* color spaces. The tests were conducted in a variety of industrial samples. The obtained results are promising and show the possibility of efficiently classifying complex industrial products based on color and texture features.