Three-dimensional brain tumor MRI segmentation is a challenging task in the field of medical image analysis. Recently, deep learning methods, particularly CNN-based architectures, have significantly improved segmentation by learning complex spatial features. However, challenges such as preserving fine details, texture variation, handling class imbalance, and ensuring generalization persist. Enhancing deep learning models with domain-specific knowledge, such as texture-aware filters, can further improve segmentation accuracy and robustness. This paper presents a novel model that incorporates Gabor convolution into a U-Net architecture to enhance texture analysis and minimize feature loss. The model processes 3D brain tumor MRIs slice by slice, utilizing multi-view inputs to preserve spatial details while maintaining a lightweight design. Furthermore, it investigates the optimal kernel sizes for the Gabor filter, marking the first study to address this crucial aspect of integrating textural analysis with deep learning techniques. Experimental results show that the proposed framework improved segmentation accuracy by using a 7×7 Gabor kernel size and achieving Dice coefficients of 89.78% for WT, 85.60% for TC, and 83.55% for ET, with a mean Dice score of 86.31%. The proposed model demonstrated consistent improvements over the standard U-Net and outperformed several existing state-of-the-art methods.

The creation of efficient network intrusion detection systems (NIDS) has become vital with the rising occurrence of network intrusions. In this research, we introduce an innovative NIDS method that integrates the strengths of convolutional neural network (CNN) and long short-term memory (LSTM) mechanisms to examine the network traffic data characteristics. We employ the UNSW-NB15 dataset, which showcases a varied distribution of patterns, including a notable imbalance between the size of the training and testing sets. Unlike conventional machine learning methods, which frequently face challenges with restricted feature sets and reduced accuracy, our proposed model addresses these shortcomings. Current models applied to this dataset generally necessitate manual feature selection and extraction, which can be less accurate, labor-intensive and time-consuming. Conversely, our model attains better performance in binary classification by harnessing the benefits of combined CNN and LSTM models. By conducting thorough experiments and evaluations with advanced deep learning models, we showcase the superiority and the efficacy of the proposed approach. The obtained results emphasize the promise of integrating CNN and LSTM in order to improve network intrusion detection.

This study introduces a new approach to diagnosing Alzheimer’s disease by combining Generative Adversarial Networks (GANs) with Vision Transformers (ViTs) to tackle the common issue of limited medical imaging data. GANs are used to generate synthetic MRI images of Alzheimer’s patients, which are combined with real images to create a richer and more varied dataset. In this approach, the GAN model— comprising a generator and discriminator—learns to produce high-quality synthetic MRI images that, together with real data, significantly expand the training set. These images are then fed into a ViT model, which, thanks to its self-attention mechanisms, excels at identifying and classifying the stages of Alzheimer’s disease. Our evaluation shows impressive results, with metrics including accuracy, F1-score, AUC, precision, and recall reaching 98.8%, 98.43%, 99.5%, 98%, and 98%, respectively— showing the effectiveness of the expanded dataset in boosting classification performance. This integrated GAN-ViT approach not only enhances diagnostic accuracy for Alzheimer’s but also sets the stage for applying similar techniques across other medical imaging domains, providing better tools for understanding disease progression and improving patient outcomes.

Breast cancer (BC) is a leading cause of cancer-related deaths in women, but early detection significantly improves survival rates. Recently, deep learning neural networks have shown potential for enhancing BC screening, identification, and classification. In this research, we propose a modified deep learning model incorporating customized architecture adjustments and advanced data augmentation techniques for early BC detection and classification from mammography images. Our model achieves a 40% increase in accuracy compared to state-of-the-art transfer learning methods on the MIAS and SA datasets, demonstrating improved robustness and precision. However, while our results are promising, the limited size and diversity of the datasets suggest that further studies are needed to validate the generalizability of the model across broader, more varied data sources.

In 1988, Fere discovered the Electrodermal activity (EDA) and it was defined originally as the property of human skins. Nowadays, it is well known as the characteristics of the human body that causes an incessant variation of the electrical skin potential. In this work, the EDA signal is used to detect the Polyneuropathy (PNP) disease. The main two steps of the proposed strategy is to extract several features via EDA signals and to classify them in two classes (Healthy case and PNP case) by using Support Vector Machine (SVM) algorithm. For this purpose, four different domains of feature extraction are investigated (morphology, time, frequency and time-frequency). The Emrirical Mode Decomposition (EMD) algorithm is used to decompose original EDA to some sub-signals ranged from high to low frequency order. Consequently, the time-frequency domain is investigated, and the EDA analyse is performed considering diffirent frequency ranges. Then, the extracted features were classified using SVM and 83.79% of accuracy was achieved. Compared to previous works, experimental results show that the proposed method is truthful for PNP detection.

Melanoma is the most lethal form of skin cancer, but early detection can lead to effective treatment. Subsequently, the main concern of the health management community is to create efficient systems to detect melanoma earlier by utilizing computer vision systems since the traditional screening methods are manual, time-consuming, and inaccurate in some cases. These systems use measurable visual components describing the shape, color, and texture. These features are extracted based on rules invented by dermatologists to determine the malignancy of skin lesions. In this paper, we propose a novel approach to melanoma detection based on the “ugly duckling” concept, which suggests that nevi in the same individual usually resemble each other, and malignant melanomas often do not follow this pattern. Our method uses a convolutional neural network architecture to extract feature vectors from dermatoscopic images of skin lesions. Then, out-liers are detected by applying principal component analysis. The outliers are indicative of potential melanoma lesions. We evaluate the performance of our method using a dataset of dermatoscopic images. Our proposed method has shown the potential to improve melanoma detection rates.

This paper introduces a method for classifying and recognizing flying objects using trajectory features and artificial neural networks (ANN). Initially, the video sequence undergoes processing through a Gaussian mixture model (GMM) to detect and track the flying objects. Then, we will extract trajectory features from our dataset and use them to feed ANN for flying object classification. These features include turning angle, speed, acceleration and centroid distance function. The classical ANN is applied for feature vector classification to discriminate between birds and drones. Experimental results are conducted to showcase the effectiveness of the proposed method in classifying drones and birds. Moreover, the automated approach can be valuable in aiding military services to differentiate drones from other objects.

Wildfire is an unplanned natural disaster that can threaten human lives. The assessment of wildfires and smoke flows is difficult and needs systems with good precision and rapidity. This study proposes an effective workflow for detecting fire and smoke in RGB video surveillance images with two basic stages: we suggest first to enhance the image quality and denoise the noisy images with a lightweight convolutional encoder decoder architecture then we detect fire and smoke with a preprocessing steps combined with an adaptive level set algorithm. The experimental results of the proposed method prove its ability and efficiency in detecting fires and smokes with a minimum of false negatives regions and over than 90% value of Jaccard index.

Early fire forest detection is crucial for fast and effective intervention. Many research have been done on this subject starting by sensor based systems and arriving to image processing which leverage the computer vision advancements. Our work refers to one of the latest algorithms in forest fire detection: YOLO. We present in this paper a detailed description of the architecture of the YOLO algorithm with an emphasis to the YOLOv6 which is the latest version of the YOLO algorithms. The performance of the studied algorithm is evaluated on a personal database containing 28334 images, with 10534 forest fire images and 17800 non-fire images. The experimental results of applying the YOLOv6 proved the efficiency of the method in fast and accurate fires detection even in large size images and low resolutions. This result makes the studied algorithm so suitable for both satellite and ground based images analysis.

In this article, we present a new method for automatic license plate recognition (ALPR) based on local power spectrum (LPS) features map and convolutional neural network (CNN). The multi-scaled and multi-oriented LPS features derived from log Gabor wavelets are well discussed. Hence, LPS at given orientation and scale is applied for license plate detection (LPD). Then, we apply an adaptive thresholding algorithm to LP character string for binarization. After that, characters are extracted separately to feed deep CNN for the Tunisian LPR. The proposed LPD approach is tested on Tunisian and Benchmark datasets under different conditions of complexities. Our developed system achieves about 96% accuracy on LPD and 95% on LPR.

Gait can be defined as the individuals manner of walking. Its analysis can provide significant information about their identity and health, opening a wide range of possibilities in the field medical diagnosis. In this paper we have introduced an innovated method that implement the artificial intelligence to classify human gait in video sequence. We exploited modern library based on deep learning to estimate human pose, extract the gait features and analyses videos, we have used conventional neural networks to classify normal and pathological gait. We have used Anaconda to code with Python and some packages such as Jupyter, Keras,Tensorflow, MediaPipe and DeepPose.

To develop an efficient device that helps dermatologists to early evaluate and inspect a specific kind of skin disease, computer vision systems have been intensively studied. These systems replace the traditional screening ways which are manual and time-consuming. These systems use some measurable visual component describing the shape, color, and texture of skin diseases to recognize them and to specify their malignancy. This article will be concentrated on the importance of using some statistical features and extracting the most relevant features of texture-colored images by calculating their degree of characterization. Using these highly-rated static textural features, non-fatal skin disease and melanoma classification results are presented and discussed.

Fires have become a more serious hazard to people's lives, property, and environment. Compared with the traditional techniques of fire detection, image technologies play a very promising role to overcome the problem of high false alarm rate. However, a major issue with these methods is their fastidious and long-time generation. In fact, the implemented algorithms are often produced using multi-feature technique, including chromatic characteristics, dynamic features, texture features and contour features. Therefore, we provide, in this paper, a study of some supervised machine learning algorithm for fire and smoke images recognition, and we compare it to a proposed model based on convolution neural network (CNN) algorithm. To do this, we consider a proper database composed by a total of 28334 images classified into three categories: 7329 fire images, 9205 smoke images and 11800 other images.

Abstract In recent decades, global warming has contributed to an increase in the number and intensity of wildfires destroying millions hectares of forest areas and causing many casualties each year. Firemen must therefore have the most effective means to prevent any wildfire from breaking out and to fight the blaze before being unable to contain and extinguish it. This article will present a new network architecture based on Convolutional Neural Network to detect and locate smoke and fire. This network generates fire and smoke masks in an RGB image by segmentation. The purpose of this work is to help firemen in assessing the extent of fire or monitor an incipient fire in real time with a camera embedded in a vehicle. To train this network, a database with the corresponding images and masks has been created. Such a database will allow to compare the performances of different networks. A comparison of this network with the best segmentation networks such as U‐Net and Yuan networks has highlighted its efficiency in terms of location accuracy, reduction of false positive classifications such as clouds or haze. This architecture is also efficient in real time.

Power electronic converters such as inverters and rectifiers are crucial parts of renewable energy systems. Usually, the power converters are subjected to a high failure frequency rate and lead to power system shut down. In an effort to predict power insulated gate bipolar transistor (IGBTs) device aging, this paper proposes a remaining useful life estimation algorithm for degraded power IGBTs, which are exposed to high amplitude of thermal overstress, through the Gaussian Process regression technique. The benefits of the proposed prognostic method are also illustrated with real device accelerated aging database set under thermal overstress utilizing a DC at the gate.

Common mechanical failures in wind turbine generators (WTGs) result in unplanned downtime, loose of production and increase the maintenance cost. Statistical studies have shown that failures due to high-speed shaft bearing (HSSB) account for 64% of all drivetrain failures. Consequently, prognostic and health management (PHM) of WTGs aims to estimate the future state of health and predict the reaming useful life (RUL) of HSSB. This paper considers a new data-driven approach based on vibration signals. This approach extracts statistical time-domain features that reflect the behavior of the system and its degradation. Then, the extracted features are evaluated to select the most trendable condition indicators that will be considered as inputs for an Elman neural network (ENN). Moreover, this paper proposes a new ENN architecture for direct RUL estimation of HSSB validated by use of real measured data from a WTG drivetrain. The proposed method reveals accurate estimation capability even with noisy measurements and harsh conditions.

The human face is viewed as the mirror that reflects the inner feelings of the person and allows us to detect the needs of every person and study his behaviour and requirements. We can also identify the tastes of each person and predict his reaction through facial features. Indeed, the detection, classification, characterization of the face are research areas that have received considerable interest in the recent twenty years. However, most published works in this field study the emotions for a person in an upright posture, which means that the person must perfectly face the camera. The innovation in the present work is to improve the detection of emotions in the case of different face orientations (to the right, left, up and down). Eight-teen feature points that perfectly characterize the human face are firstly calculated. An optimization step is proposed by extracting a set of optimal distances between the facial points as a new set of optimized emotion descriptors. For this reason, we have used a statistical characterization criterion based on the ratio of the intra-class variance. An emotion classification experiment is carried out using a multilayer neural network implemented with a digital signal processing microcontroller.

This work presents an improved procedure able to achieve straight classification of fetal abnormality in head ultrasound images in order to supply quantitative assessment of healthy or fetal hydrocephalus (H-/H+) cases. Indeed, the majority of physician rely on manual diagnostic by the use of morphological characteristics before interpreting the clinical implication of all fetal region measurements. The proposed method deals with the discriminant textural features extraction from fetal head dataset that can contribute on the recognition of cerebral diseases. The main contribution of this work is the proposal of a fully computerized approach of fetal hydrocephalus detection using relevant textural features in order to study its aptitude for evaluating the abnormal subjects within a reduced processing time. Experimental results on fetal US images show the efficiency of the proposed method when compared to the manual delineations of experts' evaluation. The proposed scheme provides suitable hydrocephalus classification rates and show a good ability to distinguish this anomaly in early stage.

Diagnosis of vertigo disorders presents many complications in the evaluation and treatment. In clinical practice, videonystagmography (VNG) tests are still an excellent bedside examination tool for vestibular disorder diagnosis. The parameters of different tests are used to get significant medical characterization of this disease. In this paper, we propose an approach to develop the assessment of vertigo symptom by the selection of the most pertinent VNG parameters using Fisher Linear Discriminant analysis. Therefore, a multilayer neural network (MNN) classifier is applied for automatic VNG dataset analysis based upon the fundamental measurements of normal and affected patients by vestibular disorder. The experimental results prove that the proposed approach is very interesting and helpful for an accurate diagnostic of this disease.

Accurate diagnostic and prognostic of fetus detects is an important challenge based on fetal head formation to supply much critical information that requires more attention in evaluating the abnormal heads. One of the fundamental problems currently faced, is how to limit the low signal to noise ratio with respect to the complexity of small fetal head ultrasound images dimension. This paper deals with a fully automatic detection system of subsequent fetal head composition from ultrasound images. In the preprocessing task, two filters have been used for speckle noise reducing. Using the Hough transform technique, fetal head structure detection is achieved, giving 97% as segmentation accuracy. Experimental results are analyzed using five ultrasound sequences that illustrate the effectiveness and the accuracy of the proposed method for a factual diagnostic of fetal heads.

The document image segmentation is an indispensable task in the document layout analysis system. This paper presents an accurate segmentation approach based on fuzzy classification for the administrative document image. The texture-based analysis works for this kind of document image are rare. And the research works on specific tasks are limited. Moreover, the texture-based segmentation methods are desired because they do not rely strongly on a priori knowledge surrounding the document. In addition, the robustness of these methods for degraded documents has been proven. For these purposes, the texture is explored in the analysis for our image type, using a fuzzy classification. The Fisher score determinate the most discriminative texture features for our segmentation: mean and variance. Our approach achieves encouraging and promising results for the detection of document zones: text, image and background. Qualitative and quantitative experiments are presented to determinate our approach performance.

Our objective of this project is to apply the theory of linear algebra called “singular value decomposition (SVD)” to digital image processing, specifically for fingerprint images verification. For optimal recognition, we proceed in two steps. In the first step, we begin by identifying the fingerprint features with SVD approach. In the second step, the classification accuracy of the proposed approach is evaluated with Back Propagation Neural Network (BPNN) classifier. I have implemented many extensive experiments, they prove that the fingerprint classification based on a novel SVD features and the BPNN give better results in fingerprint verification than several other features and methods.

Nystagmus is an involuntary oscillation of the human eye. The accurate detection of eye movement is based on horizontal and vertical component of the eye. In this paper a new component of the nystagmus is proposed for further processing saccade parameters such as velocity, amplitude and duration. For the reliability of the eye movement analysis a video-nystagmography (VNG) is deployed by the medical institutions in the world.

This paper presents a method for the accurate detection of the positions of moving phlebotomenae (sand-flies), and the extraction of their complete shape. The proposed method is based on the background subtraction approach, with a statistical background model found using Gaussian mixture model. Furthermore, a method based on the maximization of interclass variance, is used to eliminate wings of the phlebotomenae and then detect accurately its position. Results are further used to study the behaviour of phlebotomenae, and subsequently, improve the traps to fight against many diseases transmitted by these insects, especially leishmaniasis. The experimental results show the efficiency of the proposed algorithm to accurately detect the position and extract the complete shape of moving phlebotomenaes even in case of very blurry ones.

This paper present a new method to resolve the problem of the estimation of eye position in the analysis of videonystagmography(VNG) sequences, that studies eye vibration using active contour model. An algorithm for horizontal and vertical nystagmus tracking based on some parameters such as position, amplitude and duration, is presented. Indeed, the algorithm uses active contour method to segment the pupil. Thus, segmentation is obtained from Otsu's thresholding, to detect an ellipse region approximates the pupil form. Then, the snake-model uses the estimated ellipse as the initial contour. This proposed approach is approved on clinical samples of the videonystagmography (VNG) from patient who presented a congenital nystagmus.

Our objective of this project is to apply the theory of linear algebra called “singular value decomposition (SVD)” to digital image processing, specifically for fingerprint images verification. For optimal recognition, we proceed in two steps. In the first step, we begin by identifying the fingerprint features with SVD approach. In the second step, the classification accuracy of the proposed approach is evaluated with Back Propagation Neural Network (BPNN) classifier. I have implemented many extensive experiments, they prove that the fingerprint classification based on a novel SVD features and the BPNN give better results in fingerprint verification than several other features and methods.

In this paper, we propose a new method for fingerprint identification based on the Euclidian distance between the center point and their nearest neighbor bifurcation minutiae's. The main advantage of the new method is the reduced number of features vectors used to characterize fingerprint, compared with the classic characterization method based on the spatial coordinate position of bifurcation minutiae points. In addition, this new method avoids the problem of geometric rotation and translation over the acquisition phase of image fingerprints. Whatever the degree of fingerprint rotation, the extraction features used to characterize the fingerprint remains the same. The characterization efficiency of the proposed method is compared to the method based on the spatial coordinate position of fingerprint minutiae. The comparison is based on a characterization criterion, usually used to evaluate the class quantification and the features discriminating ability. Extensive experiments prove that the Fingerprint Characterization based on the Euclidean distance between the center point and their nearest neighbor bifurcation minutiae's gives better results in fingerprint classification than several other features.

We propose, in this paper, a new non-linear exponential adaptive bi-dimensional (2D) filter for image modeling. The filter coefficients are updated with the least mean square (LMS) algorithm. Furthermore, the proposed nonlinear model is used for texture modeling with a 2D auto-regressive (AR) adaptive model. The characterization efficiency of the proposed exponential model is compared with the 2D linear AR model updated with the LMS algorithm. The comparison criteria is based on the computation of a characterization rate using the ratio of "between-class" variances with respect to "within-class" variances of the estimated coefficients. Extensive experiments show that the exponential model coefficients give better results in texture discrimination than those of the linear model, even in a noisy context.

In this paper, we propose an efficient approach based on a fast convolution algorithm to reduce the computational complexity of the Least Mean Square (LMS) adaptive algorithm for the quadratic filter, i.e. the quadratic part of the second order Volterra filter (SOVF). The previous works using the fast convolution in the adaptive LMS filtering are limited to the linear case. We show that this approach reduces the multiplications number by close to 25%, at the expense of only 25% more additions. The steady-state performance of this algorithm is studied for gaussian inputs and in stationary setting. The Steady-State Excess Mean-Square-Error is evaluated. The theoretical performance predictions are shown to be in good agreement with simulation results, especially for small step-sizes.

In the framework of parametric texture modeling, a question arises: are adaptive approaches based on higher order statistics (HOS) more appropriate to characterize texture models than those based on second order statistics (SOS)? In order to give some responses to this question, we have compared two fast adaptive filters for texture characterization: the 2-D FLRLS filter (2-D fast lattice recursive least square) based on SOS only and the 2-D OLRIV filter (2-D overdetermined lattice recursive instrumental variable) based on third order statistics. Extensive experiments to study the characterization performance of each filter are presented and interpreted. They show that the 2-D FLRLS filter provides a very good performance for texture characterization, even with important noise. Furthermore, the third order based algorithm presents higher variance than the second order one. We believe that for 2-D adaptive modeling, there is no advantage to using a HOS based adaptive algorithm for characterizing textures.

We take into account the non-Gaussian properties of textures and we propose a new approach for their characterization based on bidimensional adaptive modeling using higher order statistics. The 2D-OLRIV (bidimensional overdetermined lattice recursive instrumental variable) algorithm allows accurate texture model estimation. Sets of 2D-AR coefficients obtained from the 2D reflection coefficients of the lattice model are used to characterize the texture model. This algorithm has the advantage of yielding non-biased estimates of the 2D-AR model even when the texture image is disturbed by Gaussian noise. A multilayer neural network deals with these coefficients in order to classify different textures. In order to evaluate the performance of this approach, the classification sensitivity is evaluated on a set of eight different textures. This characterization approach gives very promising results.

This paper presents a fast method for nonlinear filtering based on multichannel Chandrasekhar equations. By assuming that the adaptive second order Volterra filter may be transformed in a multichannel input linear filter, we present a new form of the second order Volterra filtering based a the fast multichannel Chandrasekhar algorithm. This method has a computational complexity of 3.N/sup 3/ multiplications per time instant, where N represents the memory span in number of samples of the nonlinear system model. This compares with 7.N/sup 3/ multiplications required for application of the fast Kalman filter with the same approach. A direct implementation of the RLS algorithm has a computational complexity of N/sup 6/. The adaptive filter is successfully used in a second order Volterra system identification in a stationary environment.