Transfer of Learning

Deep learning-based classification and detection algorithms have emerged as a powerful tool for vehicle detection in intelligent transportation systems. The limitations of the number of high-quality labeled training samples makes the single vehicle detection methods incapable of accomplishing acceptable accuracy in road vehicle detection. This paper presents detection and classification of vehicles on publicly available datasets by utilizing the YOLO-v5 architecture. This paper’s findings utilize the concept of transfer learning through fine tuning the weights of the pre-trained YOLO-v5 architecture. To employ the concept of transfer learning, extensive data sets of images and videos of the congested traffic patterns were collected by the authors. These datasets were made more comprehensive by pointing various attributes, for instance high- and low-density traffic patterns, occlusions, and different weather circumstances. All of these gathered datasets were manually annotated. Ultim...

Different diseases can affect an individual's gait in different ways and, therefore, gait analysis can provide important insights into an individual's health and well-being. Currently, most systems that perform gait analysis using 2D video are limited to simple binary classification of gait as being either normal or impaired. While some systems do perform gait classification across different pathologies, the reported results still have a considerable margin for improvement. This paper presents a novel system that performs classification of gait across different pathologies, with considerably improved results. The system computes the walking individual's silhouettes, which are computed from a 2D video sequence, and combines them into a representation known as the gait energy image (GEI), which provides robustness against silhouette segmentation errors. In this work, instead of using a set of handcrafted gait features, feature extraction is done using the VGG-19 convolutional neural network. The network is fine-tuned to automatically extract the features that best represent gait pathologies, using transfer learning. The use of transfer learning improves the classification accuracy while avoiding the need of a very large training set, as the network is pre-trained for generic image description, which also contributes to a better generalization when tested across different datasets. The proposed system performs the final classification using linear discriminant analysis (LDA). Obtained results show that the proposed system outperforms the state-of-the-art, achieving a classification accuracy of 95% on a dataset containing gait sequences affected by diplegia, hemiplegia, neuropathy and Parkinson's disease, along with normal gait sequences.

HAL is a multi-disciplinary open access archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L'archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d'enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Background and objective: Lyme disease which is one of the most common infectious vectorborne diseases manifests itself in most cases with erythema migrans (EM) skin lesions. Recent studies show that convolutional neural networks (CNNs) perform well to identify skin lesions from images. Lightweight CNN based pre-scanner applications for resource-constrained mobile devices can help users with early diagnosis of Lyme disease and prevent the transition to a severe late form thanks to appropriate antibiotic therapy. Also, resource-intensive CNN based robust computer applications can assist non-expert practitioners with an accurate diagnosis. The main objective of this study is to extensively analyze the effectiveness of CNNs for diagnosing Lyme disease from images and to find out the best CNN architectures considering resource constraints. Methods: First, we created an EM dataset with the help of expert dermatologists from Clermont-Ferrand University Hospital Center of France. Second, we benchmarked this dataset for twentythree CNN architectures customized from VGG, ResNet, DenseNet, MobileNet, Xception, NASNet, and EfficientNet architectures in terms of predictive performance, computational complexity, and statistical significance. Third, to improve the performance of the CNNs, we used custom transfer learning from ImageNet pre-trained models as well as pre-trained the CNNs with the skin lesion dataset HAM10000. Fourth, for model explainability, we utilized Gradientweighted Class Activation Mapping to visualize the regions of input that are significant to the CNNs for making predictions. Fifth, we provided guidelines for model selection based on predictive performance and computational complexity. Results: Customized ResNet50 architecture gave the best classification accuracy of 84.42% ±1.36, AUC of 0.9189±0.0115, precision of 83.1%±2.49, sensitivity of 87.93%±1.47, and specificity of 80.65%±3.59. A lightweight model customized from EfficientNetB0 also performed well with an accuracy of 83.13%±1.2, AUC of 0.9094±0.0129, precision of 82.83%±1.75, sensitivity of 85.21% ±3.91, and specificity of 80.89%±2.95. All the trained models are publicly available at , which can be used by others for transfer learning and building pre-scanners for Lyme disease. Conclusion: Our study confirmed the effectiveness of even some lightweight CNNs for building Lyme disease pre-scanner mobile applications to assist people with an initial self-assessment and referring them to expert dermatologist for further diagnosis.

Denoising speech signals represent a challenging task due to the increasing number of applications and technologies currently implemented in communication and portable devices. In those applications, challenging environmental conditions such as background noise, reverberation, and other sound artifacts can affect the quality of the signals. As a result, it also impacts the systems for speech recognition, speaker identification, and sound source localization, among many others. For denoising the speech signals degraded with the many kinds and possibly different levels of noise, several algorithms have been proposed during the past decades, with recent proposals based on deep learning presented as state-of-the-art, in particular those based on Long Short-Term Memory Networks (LSTM and Bidirectional-LSMT). In this work, a comparative study on different transfer learning strategies for reducing training time and increase the effectiveness of this kind of network is presented. The reduction in training time is one of the most critical challenges due to the high computational cost of training LSTM and BLSTM. Those strategies arose from the different options to initialize the networks, using clean or noisy information of several types. Results show the convenience of transferring information from a single case of denoising network to the rest, with a significant reduction in training time and denoising capabilities of the BLSTM networks.

The quality of speech signals is affected by a combination of background noise, reverberation, and other distortions in real-life environments. The processing of such signals presents important challenges for tasks such as voice or speaker recognition. To enhance signals in such challenging conditions several deep learning-based methods have been proposed. Those new methods have proven to be effective, in comparison to classical algorithms based on statistical analysis and signal processing. In particular, recurrent neural networks, especially those with long shortterm memory (LSTM and BLSTM), have presented surprising results in tasks related to enhancing speech. One of the most challenging aspects of artificial neural networks is to reduce the high computational cost of the training procedure. In this work, we present a comparative study on transfer learning to accelerate and improve traditional training based on random initialization of the internal weights of the networks. The results show the advantage of the proposal in terms of less training time and better results for the task of denoising speech signals at several signalto-noise ratio levels of white noise.

Transfer is the ability to employ knowledge acquired in one task to improve performance in another. We study transfer in the context of the ICARUS cognitive architec-ture, which supplies diverse capabilities for execution, inference, planning, and learning. We report on an ...

Classification of brain tumors is a difficult problem in medical imaging analysis. Over the past few years, various deep learning-based techniques have been employed for detecting
and classifying tumors from Computed Tomography (CT) and Magnetic Resonance Imaging
(MRI). In this paper, a new model referred to as Multimodal Attention based Convolutional
Neural Network (MANet) has been proposed for detection and classification of brain tumors.
The proposed MANet architecture involves integrating multiple feature sets extracted from
MRI images through attention techniques. These include wavelet features, edge features,
and texture features. This network is divided into two modules, namely feature extraction
and classification. The feature extraction module incorporates three CNN streams assisted by wavelet, edge, and texture attention, respectively. The three streams learn discriminative
features from diverse feature sets. The classification module concatenates the features and
makes predictions through fully connected layers. The proposed MANet models four classes
of brain tumors. These classes include glioma, meningioma, pituitary, and non-tumor classes.
The proposed MANet model has been tested using two publicly available datasets. These datasets include Contrast Enhanced Magnetic Resonance Imaging (CT-MRI) and Brain
Tumor Magnetic Resonance Imaging (BT-MRI). Comparison has been made between MANet
and several other existing methods. Experimental results show that the proposed MANet
model outperforms other approaches. Comparisons have been made to many existing state-
of-the-art approaches. The experimental findings suggest that the suggested MANet is more
effective than other approaches. MANet model reaches the highest level of efficiency with
an accuracy of 99.12%, a precision of 99.44%, a recall of 99.35%, a sensitivity of 99.1%, a
specificity of 99.6%, and an F1 score of 99.5%. Furthermore, experiments show the efficacy of different attention mechanisms along with suggested multimodal fusion technique. In conclusion, the proposed MANet is very promising as a tool to classify brain tumours.

Speech emotion recognition (SER) is a challenging field of research that has attracted research during the last two decades. Successful performance of Deep Convolutional Neural Networks (DNNs) in various difficult pattern recognition problems motivates researchers to develop SER systems based on deep learning algorithms. The most essential requirement in training a deep model is the presence of a large-scale dataset. However, in many cases, such an amount of data is not available. Transfer learning approaches provide a practical solution to this problem. In this paper, we proposed an SER system based on AlexNet, the well-known deep model trained on the large-scale ImageNet dataset. In this way, the novel enriched spectrogram calculated based on the fusion of wide-band and narrow-band spectrograms is developed as a proper input for such a model. The proposed fused spectrogram benefited from both high temporal and spectral resolution. These images have been applied to the pre-trained ...

The spectral analysis and spatial analysis of high dimensional images are very important and in this paper we tried to cover some aspects that how this problem can be handled and proposed a way through which we can overcome the problem of the time constraint and using some deep learning novel approach like transfer learning for getting the best results while performing the actual computations and the results which we obtained. The dataset used is EuroSAT in which by using a VGG network, the accuracy is achieved 95 per cent and the validation accuracy achieved is 92 per cent. Also, the Kappa score which we got for this observation is 0.95. The tool used for the implementation purpose is TensorFlow with GPU which is also discussed in the paper.

Early and accurate identification of rice leaf diseases is essential for sustainable crop management; however, many existing convolutional neural networks (CNNs) based solutions struggle with class imbalance and limited robustness when applied to real-field data. In this work, a generative adversarial network (GAN) augmented vision transformer (ViT) framework is introduced to overcome these limitations. A deep size representative samples for underrepresented disease categories, resulting in a more balanced training dataset and achieving a Fréchet inception distance (FID) score of 18.6. The balanced dataset is then used to train a vision transformer model that leverages self-attention to capture global contextual features of rice leaf images. Experimental evaluation across ten disease classes shows that the proposed approach attains an overall classification accuracy of 96.5%, exceeding the performance of several established CNN architectures. Additionally, the model demonstrates strong generalization capability on an external field dataset, achieving 94.8% accuracy. To validate real-world applicability, the trained model is deployed on a Jetson Nano edge device, where it delivers efficient inference performance suitable for practical agricultural applications. The findings indicate that combining GAN-based data augmentation with transformer-based learning provides a reliable and scalable solution for rice leaf disease detection.

Facial expressions play a key role in non-verbal communication by naturally reflecting human emotions. Facial emotion recognition (FER) using computer vision has gained attention with advances in deep learning. However, deep learning models require large datasets to perform well, posing a challenge for FER tasks with limited data. Transfer learning is a promising approach to address this issue, but a standardized method for FER is yet to be established. This study optimizes three transfer learning models ResNet-50, Inception V3, and Xception on the FER-2013 dataset. Experiments include testing input image sizes, hyperparameter tuning, data augmentation, layer addition, and training methods. Results show each model requires different input sizes for best accuracy. Hyperparameter tuning improves accuracy by 6.35%, 4.69%, and 1.04% for ResNet-50, Inception V3, and Xception, respectively. Augmenting only the disgust class yields better accuracy than augmenting all classes. The freeze fine-tuning method is less effective than fine-tuning alone on datasets with thousands of samples but outperforms the freeze layer method. The best accuracies achieved are 64.89% (ResNet-50), 65.83% (Xception), and 66.40% (Inception V3). These findings provide insights into freeze fine-tuning limitations and guidance for optimizing transfer learning in FER with limited data.

first occurred in Wuhan, China in December 2019. Subsequently, the virus spread throughout the world and as of June 2020 the total number of confirmed cases are above 4.7 million with over 315,000 deaths. Machine learning algorithms built on radiography images can be used as a decision support mechanism to aid radiologists to speed up the diagnostic process. The aim of this work is to conduct a critical analysis to investigate the applicability of convolutional neural networks (CNNs) for the purpose of COVID-19 detection in chest X-ray images and highlight the issues of using CNN directly on the whole image. To accomplish this task, we use 12-off-the-shelf CNN architectures in transfer learning mode on 3 publicly available chest X-ray databases together with proposing a shallow CNN architecture in which we train it from scratch. Chest X-ray images are fed into CNN models without any preprocessing to replicate researches used chest X-rays in this manner. Then a qualitative investigation performed to inspect the decisions made by CNNs using a technique known as class activation maps (CAM). Using CAMs, one can map the activations contributed to the decision of CNNs back to the original image to visualize the most discriminating region(s) on the input image. We conclude that CNN decisions should not be taken into consideration, despite their high classification accuracy, until clinicians can visually inspect and approve the region(s) of the input image used by CNNs that lead to its prediction.

The Covid-19 first occurs in Wuhan, China in December 2019. After that the virus spread all around the world and at the time of writing this paper the total number of confirmed cases are above 4.7 million with over 315000 deaths. Machine learning algorithms built on radiography images can be used as a decision support mechanism to aid radiologists to speed up the diagnostic process. The aim of this work is to conduct a critical analysis to investigate the applicability of convolutional neural networks (CNNs) for the purpose of COVID-19 detection in chest X-ray images and highlight the issues of using CNN directly on the whole image. To achieve this task, we first use 12-off-the-shelf CNN architectures in transfer learning mode on 3 publicly available chest X-ray databases together with proposing a shallow CNN architecture in which we train it from scratch. Chest X-ray images fed into CNN models without any preprocessing to follow the many of researches using chest X-rays in this man...

The rapid advancement of Artificial Intelligence (AI), particularly in Deep Learning and Large Language Models (LLMs), has resulted in significant improvements in computational performance across domains such as natural language processing, healthcare, finance, and autonomous systems [1], [2]. However, the continuous increase in model size, typically measured by the number of trainable parameters, has led to a substantial rise in computational requirements and energy consumption [4], [5]. Modern transformer-based architectures, including large-scale models such as the GPT series, consist of billions of parameters and require extensive computational resources for both training and inference processes [1], [2].This increase in computational demand has raised serious concerns related to operational costs, scalability, and environmental sustainability due to higher electricity usage and carbon emissions from data centers [4], [6]. Studies have shown that training large deep learning models can produce significant carbon emissions, highlighting the environmental impact of large-scale AI systems [4]. This study presents a comparative analysis of small, medium, and large models to examine the impact of model size on energy consumption across both training and inference phases. The findings reveal a clear trade-off between performance and energy efficiency, where larger models often provide only marginal improvements in accuracy while consuming disproportionately higher energy [5], [7].Furthermore, the study explores energyefficient techniques such as model compression, optimized architectures, and Green AI practices to reduce computational cost and energy usage [7], [8]. The results emphasize the importance of balancing model scalability with sustainability to ensure responsible AI development in the future.

Holographic microwave imaging (HMI) has been proposed for early breast cancer diagnosis. Automatically classifying benign and malignant tumors in microwave images is challenging. Convolutional neural networks (CNN) have demonstrated excellent image classification and tumor detection performance. This study investigates the feasibility of using the CNN architecture to identify and classify HMI images. A modified AlexNet with transfer learning was investigated to automatically identify, classify, and quantify four and five different HMI breast images. Various pre-trained networks, including ResNet18, GoogLeNet, ResNet101, VGG19, ResNet50, DenseNet201, SqueezeNet, Inception v3, AlexNet, and Inception-ResNet-v2, were investigated to evaluate the proposed network. The proposed network achieved high classification accuracy using small training datasets (966 images) and fast training times.

Paddy (Rice) leaf illnesses form a notable threat to crop productivity and food efficiency, especially in developing countries where rice is regarded as a staple crop. Conventional illness diagnosis approaches rely on manual observation, which is laborious, erroneous, and unfit for large-scale farming. Deep learning methods, specifically Convolutional Neural Networks (CNNs), have remarkably depicted great success in computerized disease classification. Though, numerous existing methods are constrained by lofty computational complexity, high memory exploitation, and poor interpretability. This research offers a baseline study using MobileNetV2 for rice leaves disease detection. The proposed baseline model uses transfer learning for classification of four different rice leaf ailments: Brown-Spot, Bacterial Leaf Blight, Rice Blast, and Rice Tungro. The dataset comprises 5952 images obtained from Mendeley database and preprocessing techniques were applied on them. Experiments were carried out by using Google Colab CPU runtime under common parameters and settings. The MobileNetV2 baseline architecture has excellently produced accuracy of 99.27%, macro F1-score of 0.9929, with reasonably small model size and low latency against memory consuming CNN architectures. The results present MobileNetV2 as a reliable and strong lightweight baseline for rice leaf disease detection and offer a benchmark for future hybrid interpretable architectures.

Urban tree classification enables informed decision-making processes in urban planning and management. This paper introduces a novel data reformation method, pseudo tree crown (PTC), which enhances the feature difference in the input layer and results in the improvement of the accuracy and efficiency of urban tree classification by utilizing artificial intelligence (AI) techniques. The study involved a comparative analysis of the performance of various machine learning (ML) classifiers. The results revealed a significant enhancement in classification accuracy, with an improvement exceeding 10% observed when high spatial resolution imagery captured by an unmanned aerial vehicle (UAV) was utilized. Furthermore, the study found an impressive average classification accuracy of 93% achieved by a classifier built on the PyTorch framework, with ResNet50 leveraged as its convolutional neural network layer. These findings underscore the potential of AI-driven approaches in advancing urban tree classification methodologies for enhanced urban planning and management practices.

Melanoma, recognized as the most life-threatening form of skin cancer, poses a significant threat to life expectancy. The timely identification of melanoma plays a crucial role in mitigating the morbidity and mortality associated with skin cancer. Dermoscopic images, acquired through advanced dermoscopic tools, serve as vital resources for the early detection of skin cancer. Hence, there is an urgent need to develop a reliable and accurate Computer-Aided Diagnosis (CAD) system capable of autonomously discerning skin cancer. This study focuses on the meticulous construction of diverse skin cancer classification models, specifically employing various Convolutional Neural Network (CNN) architectures configured across four distinct layer arrangements. Additionally, a transfer learning approach is explored, leveraging robust pre-trained deep CNN models extensively trained on the comprehensive ISIC dermoscopic image dataset, known for its diversity in skin lesions. Utilizing the ISIC dataset as the foundation of our analysis, the CNN model's performance is systematically evaluated with varying numbers of layers-ranging from 15 to 27. Results indicate that the CNN model comprising 15 layers achieves an accuracy of 89.55%, while the model with 27 layers exhibits the highest performance, attaining an accuracy of 90.85%. In the realm of transfer learning, ten baseline CNN models pre-trained on ImageNet are employed. All baseline models demonstrate accuracies surpassing 80%, with SqueezeNet recording the lowest accuracy at 80.89%. In contrast, the ResNet-50 model consistently outperforms other models in transfer learning, achieving an accuracy of 92.98%. These findings underscore the efficacy of the proposed models in melanoma classification and highlight the superior performance of the ResNet-50 model in the context of transfer learning.

This review synthesizes on age and gender prediction from facial images, focusing exclusively on methodologies leveraging artificial intelligence (AI), machine learning (ML), and deep learning (DL). The purpose is to evaluate the evolution and efficacy of these techniques in addressing challenges such as unconstrained imaging conditions, dataset variability, and real-time applicability. AI approaches provide foundational frameworks for feature extraction and classification, while ML methods emphasize traditional classifiers like support vector machines (SVM) and neural networks for robust pattern recognition. DL, particularly convolutional neural networks (CNNs), dominates recent advancements, incorporating transfer learning, multi-task learning, and optimization strategies to achieve superior accuracies. Key findings reveal DL models, such as VGG-16, ResNet-50, and GoogLeNet, yielding gender classification accuracies up to 99.3% and mean absolute errors (MAE) as low as 2.68 years for age estimation on datasets like MORPH-II, Adience, and UTKFace. ML techniques, including SVM and artificial neural networks (ANN), offer accuracies around 83-96% but struggle with largescale variations. Common datasets include Adience for unconstrained scenarios and MORPH-II for longitudinal aging data. Preprocessing steps like histogram equalization and data augmentation enhance performance, with evaluation metrics favoring accuracy, F1-score, and MAE. Despite progress, research gaps persist in handling occlusions (e.g., masks), ethnic diversity, and computational efficiency for mobile deployment. This review underscores DL's transformative role while highlighting the complementary strengths of ML for resource-constrained environments.

Surface defects in solar panels have a detrimental impact on the performance and reliability of the system. In this research paper, we proposed an automatic fault detection model for solar panels using a deep learning model based on the VGG16 (CNN) architecture with transfer learning and data augmentation. A dataset that has a total of 869 images for six fault categories (Clean, Dusty, Bird-drop, Electrical-damage, Physical-damage, Snow) was used. Two models were utilized for analysis: (Model 1) the baseline VGG16 which is the VGG16 architecture pre-trained on the original dataset scope; and (Model 2) the enhanced VGG16 which incorporates data augmentation, freezing of early layers and dropout regularizing. Model 2 achieved a higher accuracy at 98.76% on validation set compared to Model 1 with an accuracy of 82.49% indicating that comprising transfer learning along and pertinent augmentation positively enhances the classification accuracy. The proposed approach is capable of being scaled up and replicated in real-world applications of smart automated solar panel inspection.

This work focuses on the automatic classification of deontic sentences. It presents a novel Machine Learning approach which combines the power of Transfer Learning with the information provided by two famous LegalXML formats. In particular, different BERT-like neural architectures have been fine-tuned on the downstream task of classifying rules from the European General Data Protection Regulation (GDPR) encoded in Akoma Ntoso and LegalRuleML. This work shows that fine-tuned language models can leverage the information provided in LegalXML documents to achieve automatic classification of deontic sentences and rules.

Some recent studies have described deep convolutional neural networks to diagnose breast cancer in mammograms with similar or even superior performance to that of human experts. One of the best techniques does two transfer learnings: the first uses a model trained on natural images to create a ''patch classifier'' that categorizes small subimages; the second uses the patch classifier to scan the whole mammogram and create the ''single-view whole-image classifier''. We propose to make a third transfer learning to obtain a ''two-view classifier'' to use the two mammographic views: bilateral craniocaudal and mediolateral oblique. We use EfficientNet as the basis of our model. We ''end-to-end'' train the entire system using CBIS-DDSM dataset. To ensure statistical robustness, we test our system twice using: (a) 5-fold cross validation; and (b) the original training/test division of the dataset. Our technique reached an AUC of 0.9344 using 5-fold cross validation (accuracy, sensitivity and specificity are 85.13% at the equal error rate point of ROC). Using the original dataset division, our technique achieved an AUC of 0.8483, as far as we know the highest reported AUC for this problem, although the subtle differences in the testing conditions of each work do not allow for an accurate comparison. The inference code and model are available at INDEX TERMS Breast cancer diagnosis, deep learning, convolutional neural network, mammogram, transfer learning.

In a cross-sector federated learning setup where hospitals and banks jointly train a cybersecurity AI, a quiet but dangerous problem emerges: hospitals often take weeks to report and remediate cyber threats, while banks must act within days or hours. This timing mismatch creates a subtle data leak, delayed model updates from healthcare nodes can unintentionally reveal patterns that expose sensitive banking data to an adversary. This paper introduces a human-centered generative AI framework that simulates and blocks such-time-based data spills.‖ Using a time-conditioned generative adversarial network (GAN), we first mimic how delayed healthcare updates distort the shared learning process. Then, a second defensive AI. a differentially private federated learner with adaptive noise learns to mask sector-specific footprints without forcing banks to wait for slow partners. We test the framework on an AWS cloud testbed using synthetic but realistic healthcare and banking transaction logs. Results show that our approach reduces cross-sector information leakage by over 70% while preserving detection accuracy above 90% for both sectors. Instead of demanding perfect synchronization, this work embraces real-world regulatory delays and offers a practical, privacy-respecting path for collaborative cybersecurity across industries that do not march to the same clock.

With the spread of the COVID19 pandemic, blended learning has become one of the most used methods in educational organizations such as universities, community colleges, and schools. In blended learning, the students' practical activities are done in more than one way, including simulation software and the place of study. For chemical experiment programs, the classification of handwritten chemical formulas plays an important role in determining the simulation software's efficiency. Accordingly, in this study, we propose a model for handwritten chemical formula classification. First, this paper describes a handwritten chemical formulas dataset that contains eight classes (HCFD8). Second, convolutional neural networks (CNNs) with pre-trained weights are used as a deep feature extractor to extract features from the images. Third, due to limited training images per class, the proposed model uses data augmentation techniques to expand the training images. Then, an enhanced multilayer perceptron (EMLP) strategy is used to classify the image. Finally, we provide a performance analysis of typical deep learning approaches on HCFD8, which shows that the proposed model performs good accuracy results.

The COVID-19 has become a pressing public health concern recently due to its dramatic impact. It spreads quickly, and it is beyond the ability of health staff to detect patients with the disease immediately. However, the ability to diagnose SARS-CoV-2 in a short time is critical for fighting the disease. The primary objective of this study is to develop deep neural networks to diagnose disease in a quick, safe, and cheap way. We classify the cases as normal, COVID-19, and pneumonia. Deep neural networks are developed to perform a three-class classification task. Ten deep learning models are evaluated on a large dataset. Although all DCNNs demonstrated promising potential for classification, hybrid neural networks delivered the most promising outcome with the highest accuracies. The first hybrid model is named MICOVID. The second hybrid model is named VVCOVID. These models are developed through transfer learning by using pre-trained deep learning models. Performance metrics results showed that MICOVID and VVCOVID models have an accuracy of 94% for COVID-19 detection. This is higher than other classification models. These findings suggest that two novel hybrid models that we proposed have great potential to be embedded into computer-aided systems to predict disease in radiology departments.

The COVID-19 has become a pressing public health concern recently due to its dramatic impact. It spreads quickly, and it is beyond the ability of health staff to detect patients with the disease immediately. However, the ability to diagnose SARS-CoV-2 in a short time is critical for fighting the disease. The primary objective of this study is to develop deep neural networks to diagnose disease in a quick, safe, and cheap way. We classify the cases as normal, COVID-19, and pneumonia. Deep neural networks are developed to perform a three-class classification task. Ten deep learning models are evaluated on a large dataset. Although all DCNNs demonstrated promising potential for classification, hybrid neural networks delivered the most promising outcome with the highest accuracies. The first hybrid model is named MICOVID. The second hybrid model is named VVCOVID. These models are developed through transfer learning by using pre-trained deep learning models. Performance metrics results showed that MICOVID and VVCOVID models have an accuracy of 94% for COVID-19 detection. This is higher than other classification models. These findings suggest that two novel hybrid models that we proposed have great potential to be embedded into computer-aided systems to predict disease in radiology departments.

Real-time multilingual interaction during mobile video calls still difficult to achieve due to strict latency, fluctuating network conditions, and the limited resources capacity of handheld devices. Although recent speech translation systems show high linguistic accuracy, many studies highly depend on cloud-based pipelines or text-only captions, which disrupt conversational flow and increase deployment cost. We design and implement a lightweight speech-to-speech translation system integrated directly into live video calls. Our system combines ondevice speech recognition and speech synthesis with a lightweight online translation service, coordinated via parallel pipeline that operates independently of the core audio-video stream. To avoid degradation of call quality, translated content is exchanged using a dedicated real-time data channel rather than the audio channel itself. Experiments conducted on physical Android smartphones under realistic network conditions over Wi-Fi and mobile data networks show end-to-end translation latency of approximately 0.5 to 2.0 seconds across multiple language pairs. Subjective user evaluations yield a mean opinion score of 4.2, indicating that the translated speech is sufficiently clear and natural for everyday conversational use.

The present paper aims to train and analyze Convolutional Neural Networks (CNN or ConvNets) capable of classifying plant species of a certain region for applications in an environmental monitoring system. In order to achieve this for a limited training dataset, the samples were expanded with the use of a data generator algorithm. Next, transfer learning and fine tuning methods were applied with pre-trained networks. With the purpose of choosing the best layers to be transferred, a statistical dispersion method was proposed. Through a distributed training method, the training speed and performance for the CNN in CPUs was improved. After tuning the parameters of interest in the resulting network by the cross-validation method, the learning capacity of the network was verified. The obtained results indicate an accuracy of about 97%, which was acquired transferring the pre-trained first seven convolutional layers of the VGG-16 network to a new sixteen-layer convolutional network in which the final training was performed. This represents an improvement over the state of the art, which had an accuracy of 91% on the same dataset.

Manual inspection of photovoltaic systems is expensive, hazardous, and prone to inconsistency. This paper presents SunScout, a mobile application for offline solar-panel image management and on-device fault classification. The current mobile release organizes drone, gallery, and camera captures into reusable datasets and analyzes each stored asset with a fine-tuned EfficientNetB0 classifier deployed through ONNX Runtime. On the cleaned 1,575-image dataset, a frozen MobileNetV2 baseline reached 90.79% validation accuracy, while the proposed EfficientNetB0 model achieved 94.29%, a 3.50 percentage point improvement. Together, these results show that accurate solar fault analysis can be delivered on a consumer smartphone without requiring a persistent network connection.

Accurate leaf segmentation plays a crucial role in plant phenotyping and classification tasks, as it directly influences the reliability of downstream analyses. To effectively identify target leaves in images containing distracting background elements, this study presents an integrated end-to-end fine-tuning free data processing pipeline. The proposed pipeline consists of several stages, including depth map generation, depth-aware segmentation, filtering of distant objects based on depth information, merging overlapping masks, and selecting the mask closest to the center of the input image. The resulting segmented leaf regions are subsequently used for plant disease classification with widely adopted deep learning architectures. Experimental results on in-domain test datasets indicate that the overall classification performance remains relatively stable with and without the application of the pipeline. However, under zero-shot evaluation settings, where the models are tested on previously unseen datasets, the proposed pipeline provides substantial improvements in generalization capability. Additional feature-space and distribution-level analyses further demonstrate that the proposed pipeline reduces dataset-specific bias and improves feature alignment across domains. Overall, the proposed pipeline enables accurate and efficient extraction of leaf regions even in visually complex environments, while simultaneously acting as an implicit feature regularization mechanism.

Breast cancer is one of the most dangerous types of cancer in the world among females. In the medical industry, the early detection of a breast abnormality in a mammogram can significantly decrease the death rate caused by breast cancer. Therefore, researchers directed their focus and efforts to find better solutions. Whereas researchers earlier used semi-automatic algorithms of machine learning, recently the attention is redirected toward deep learning algorithms that automatically extract features. Therefore, in the research study, two pre-trained Convolutional Neural Network models, VGG16 and ResNet50, have been used and applied on mammogram images to classify their abnormalities in terms of (1) the Benign Calcification, (2) the Malignant Calcification, (3) the Benign Mass, and (4) the Malignant Mass. The mammographic images of the CBIS-DDSM dataset are used. In the training phase, various experiments are performed on ROI images to decide on the best model configuration and fine-tuning depth. The experimental results showed that the VGG16 model provided a remarkable advancement over the ResNet50 model; the accuracy obtained was 80.0% in the first model whereas the second model could classify with a 60.0% accuracy almost randomly. Apart from accuracy, the other performance metrics used in this study are precision, recall, F1-Score and AUC. Our evaluation, based on these performance metrics, shows that accurate detection effect is obtained from the two networks with VGG16 being the most accurate. Finally, a decision support tool is developed which classifies the full mammogram images based on the fine-tuned VGG16 architecture into Benign Calcification, Malignant Calcification, Benign Mass, and Malignant Mass.

Aquatic insects and other benthic macroinvertebrates are mostly used as bioindicators of the ecological status of freshwaters. However, an expensive and time-consuming process of species identification represents one of the key obstacles for reliable biomonitoring of aquatic ecosystems. In this paper, we proposed a deep learning (DL) based method for species identification that we evaluated on several available public datasets (FIN-Benthic, STONEFLY9, and EPT29) along with our Chironomidae dataset (CHIRO10). The proposed method relies on three DL techniques used to improve robustness when training is done on a relatively small dataset: transfer learning, data augmentation, and feature dropout. We applied transfer learning by employing ResNet-50 deep convolutional neural network (CNN) pretrained on ImageNet 2012 dataset. The results show significant improvement compared to original contributions and confirms that there is a considerable gain when there are multiple images per specimen.

This study presents the development of a multimodal deep learning model for emotion recognition using video, audio, and text data extracted from a video data. The proposed system employs a structured pipeline that begins with video input, from which audio is extracted and transcribed into text using the Whisper model. Emotion recognition is performed across three modalities: facial expressions analysed using DeepFace, vocal features classified using a Long Short-Term Memory (LSTM) network, and textual features processed using a Feedforward Neural Network (FNN). Results are visualized in bar chart plotted using Seaborn and integrated into an interactive user interface built with Flet. This multimodal approach enhances accuracy and robustness over unimodal systems by leveraging complementary information from each data stream. The model has broad applications in affective computing, human-computer interaction, surveillance, and sentiment-aware technologies.

In a landscape dominated by large language models, Maghrebi Arabic dialects, though widely used in everyday communication and informal writing, remain largely underserved by Natural Language Processing (NLP) technologies. Their limited linguistic resources, high variability, and lack of standardized orthography make them particularly challenging to model effectively. To address these issues, this work introduces MagBERT, a lightweight variant of BERT designed specifically for the three major Maghrebi dialects: Algerian, Moroccan, and Tunisian Arabic, in both Arabic and Latin scripts. The model was pretrained then fine-tuned on multiple downstream tasks, demonstrating competitive performance compared to several strong benchmark models. Despite its compact size, MagBERT shows strong potential as an efficient and versatile model for processing under-resourced North African dialects.

Lung diseases continue to impose a significant healthcare burden in the United States, making early and accurate diagnosis essential for effective treatment and patient management. This study proposes an explainable fusion-based transfer learning framework for automated lung disease classification using chest X-ray and CT scan images. The proposed system integrates state-of-the-art deep learning architectures, including ResNet50, ResNet101, and EfficientNetB0, along with feature-level and decision-level fusion strategies to enhance classification accuracy and robustness. Experimental results demonstrate that the hybrid fusion models outperform standalone architectures, achieving 95.68% accuracy for chest X-ray classification and 99.61% accuracy for CT scan classification. To improve transparency and clinical reliability, Explainable Artificial Intelligence (XAI) methods such as SHAP, LIME, and Integrated Gradients were incorporated to visualize diagnostically relevant regions. The proposed framework offers an accurate, interpretable, and reliable AI-driven solution for pulmonary disease diagnosis in modern U.S. healthcare systems.

Failure to diagnose and treat retinal illnesses on time might lead to irreversible blindness. The focus is on three common retinal lesions associated with diabetic retinopathy (DR): microaneurysms (MAs), haemorrhages, and exudates. The proposed solution leverages deep learning, employing a customized residual network (ResNet) based classifier trained on real-time retinal images meticulously annotated and graded by ophthalmologists. Annotation noise was a significant obstacle addressed by downsampling and augmenting the data. Compared to cutting-edge techniques, this one performs better with test-set accuracy of 93.34% across all classes. This approach holds great promise for enhancing early detection and treatment of DR by automating the recognition of these vital retinal abnormalities. The ability to automatically classify these symptoms can aid clinicians in making more precise diagnosis and starting treatments sooner. This research shows that deep learning-based approaches are highly effective, especially when combined with a customised ResNet-based classifier and thorough pre processing steps. We observed that this method provides the ability to better the lives of patients and lower the rate of permanent blindness resulting from retinal disorders.

Skin problems are among the most common ailments on Earth. Despite its popularity, assessing it is not easy because of the complexities in skin tones, hair colors, and hairstyles. Skin disorders provide a significant public health risk across the globe. They become dangerous when they enter the invasive phase. Dermatological illnesses are a significant concern for the medical community. Because of increased pollution and poor diet, the number of individuals with skin disorders is on the rise at an alarming rate. People often overlook the early signs of skin illness. The current approach for diagnosing and treating skin conditions relies on a biopsy process examined and administered by physicians. Human assessment can be avoided with a hybrid technique, thus providing hopeful findings on time. Approaches to a thorough investigation indicate that deep learning methods might be used to construct frameworks capable of identifying diverse skin conditions. Skin and non-skin tissue must be distinguished to detect skin diseases. This research developed a skin disease classification system using MobileNetV2 and LSTM. For this system, accuracy in skin disease forecasting is the primary aim while ensuring excellent efficiency in storing complete state information for exact forecasts.

Document image classification is a challenging task due to the complexity of information contained within documents, including text, images, and their spatial arrangement. Deep learning has become a pivotal tool for extracting and learning complex patterns. However, conventional methods often grapple with integrating different data modalities and minimizing redundancy, leading to a need for more advanced and efficient deep learning strategies. This study presents a new approach to document image classification, named graph attention-driven with dual tune learning (GAD-DTL). GAD-DTL employs dual-tune learning and graph attention networks. The methodology creates semantic region embedding within document images, which incorporate both textual and spatial data. A key feature of this approach is the adaptive fusion layer, which integrates different modalities and uses a graph attention layer to capture context within each region. To minimize redundancy in learned features, we implement two distinct learning techniques, relational and non relational learning. This approach enhances document image classification by ensuring invariant representation and minimal redundancy in features.

Transfer learning and domain adaptation are critical for deploying robust artificial intelligence systems across varying environments. Machine learning models trained on a specific source distribution frequently experience performance degradation when exposed to related but distinct target distributions. This survey provides a comprehensive review of methodologies designed to mitigate this domain shift. We categorize the landscape into discrepancy-based alignments, adversarial learning strategies, graph-based and online adaptation, and parameter-efficient transfer for foundation models. Specifically, we highlight the foundational formulations established by recent researchers. Through this analysis, we provide an overview of the field's evolution from traditional statistical alignment to zero-shot and real-time adaptation in resource-constrained scenarios.

The applications of AI in the healthcare sector are increasing day by day. The application of convolutional neural network (CNN) and mask-region-based CNN (Mask-RCCN) to the medical domain has really revolutionized medical image analysis. CNNs have been prominently used for identification, classification, and feature extraction tasks, and they have delivered a great performance at these tasks. In our study, we propose a lightweight CNN, which requires less time to train, for identifying malaria parasitic red blood cells and distinguishing them from healthy red blood cells. To compare the accuracy of our model, we used transfer learning on two models, namely, the VGG-19 and the Inception v3. We train our model in three different configurations depending on the proportion of data being fed to the model for training. For all three configurations, our proposed model is able to achieve an accuracy of around 96%, which is higher than both the other models that we trained for the same thre...

Alzheimer's disease (AD) is an irreversible, progressive neurological disorder that causes memory and thinking skill loss. Many different methods and algorithms have been applied to extract patterns from neuroimaging data in order to distinguish different stages of Alzheimer's disease (AD). However, the similarity of the brain patterns in older adults and in different stages makes the classification of different stages a challenge for researchers. In this paper, convolutional neuronal network architecture AlexNet was applied to fMRI datasets to classify different stages of the disease. We classified five different stages of Alzheimer's using a deep learning algorithm. The method successfully classified normal healthy control (NC), significant memory concern (SMC), early mild cognitive impair (EMCI), late cognitive mild impair (LMCI), and Alzheimer's disease (AD). The model was implemented using GPU high performance computing. Before applying any classification, the fMRI data were strictly preprocessed. Then, low to high level features were extracted and learned using the AlexNet model. Our experiments show significant improvement in classification. The average accuracy of the model was 97.63%. We then tested our model on test datasets to evaluate the accuracy of the model per class, obtaining an accuracy of 94.97% for AD, 95.64% for EMCI, 95.89% for LMCI, 98.34% for NC, and 94.55% for SMC.

Deepfake technology has been widely adopted and has changed the face of digital media, offering unprecedented opportunities for creative expression and entertainment while simultaneously posing a significant threat to information integrity. As synthetic media becomes increasingly indistinguishable from authentic content, the risk of misinformation and digital forgery has grown, necessitating more sophisticated detection frameworks. Deep convolutional neural networks have revolutionized facial analysis, yet achieving a balance between high-precision classification and computational efficiency remains a significant challenge. We propose Attentive ResNet101, a specialized architecture that integrates a Convolutional Block Attention Module to refine feature localization and highlight anomalous facial artifacts. By strategically applying both channel and spatial attention mechanisms, the model prioritizes high-frequency details essential for deepfake detection. We conduct a rigorous comparative analysis against established benchmarks, including ResNet50, MobileNetV2, and EfficientNetB0. Our experimental results demonstrate that the Attentive ResNet101 outperforms all baseline models, achieving a superior accuracy of 96.53% and recall of 98.39%. While lightweight models such as MobileNetV2 prioritize computational efficiency, the proposed attentive framework provided the most robust sensitivity to complex facial features. These findings suggest that the integration of attention-based feature refinement is a vital step toward securing the digital media landscape against the evolving threat of synthetic content.

COVID-19 diagnosis in symptomatic patients is an important factor for arranging the necessary lifesaving facilities like ICU care and ventilator support. For this purpose, we designed a computer-aided diagnosis and severity detection method by using transfer learning and a back propagation neural network.To increase the learning capability, we used data augmentation. Most of the previously done works in this area concentrate on private datasets, but we used two publicly available datasets. The first section diagnose COVID-19 from the input CT image using the transfer learning of the pre-trained network ResNet-50. We used ResNet-50 and DenseNet-201 pre-trained networks for feature extraction and trained a back propagation neural network to classify it into High, Medium, and Low severity.The proposed method for COVID-19 diagnosis gave an accuracy of 98.5% compared with the state-of-the-art methods. The experimental evaluation shows that combining the ResNet-50 and DenseNet-201 feature...

Despite the recent ubiquity of large language models and their high zero-shot prompted performance across a wide range of tasks, it is still not known how well they perform on tasks which require processing of potentially idiomatic language. In particular, how well do such models perform in comparison to encoder-only models fine-tuned specifically for idiomaticity tasks? In this work, we attempt to answer this question by looking at the performance of a range of LLMs (both local and software-as-a-service models) on three idiomaticity datasets: SemEval 2022 Task 2a, FLUTE, and MAGPIE. Overall, we find that whilst these models do give competitive performance, they do not match the results of fine-tuned task-specific models, even at the largest scales (e.g. for GPT-4). Nevertheless, we do see consistent performance improvements across model scale. Additionally, we investigate prompting approaches to improve performance, and discuss the practicalities of using LLMs for these tasks.

Early identification of potato leaf diseases is crucial to maximise crop production while sustaining global food security. While deep
learning has transformed agricultural diagnostics, practical agricultural conditions frequently cause diminished effectiveness with lowresolution
(LR) images. This research presents integrated two-stage architecture to tackle these issues- a novel DRBNet model for image superresolution
followed by Mobilenet based approach for disease diagnosis. The proposed DRBNet method preserves both temporal and global
information, by incorporating multi-scale bidirectional interactions with identical and reduced features in its architecture. It accomplishes
excellent reconstruction with PSNR scores of 34.05, 32.08, & 31.09 and SSIM scores of 0.98, 0.90, & 0.86 with downscaling of ×2, ×4, and
×6, respectively. The MobileNet model accomplishes accurate potato leaf disease identification along with optimized computing efficiency and
fewer parameters by employing depthwise separable convolutions to develop a streamlined feature extraction method. This integrated approach
performs noticeably better than contemporary architectures, based on research findings. The suggested integrated approach obtains impressive
classification accuracies of 99.73%, 99.56%, and 99.37%, for ×2, ×4, & ×6 LR downscale factors. This integrated model provides a flexible,
versatile method for potato health monitoring in realistic smart agricultural situations, by maintaining exceptional precision even at lower
resolutions.