Original/Review Paper
Rozhin Mohammadizand; Razieh Rastgoo
Abstract
Sign language is a structured, non-vocal form of communication primarily used by individuals who are deaf or hard of hearing, who often face challenges interacting with non-signers. To address this, translation systems between sign and spoken language are essential, encompassing sign language recognition ...
Read More
Sign language is a structured, non-vocal form of communication primarily used by individuals who are deaf or hard of hearing, who often face challenges interacting with non-signers. To address this, translation systems between sign and spoken language are essential, encompassing sign language recognition and production. In this work, we focus on sign language production and propose a deep learning framework for generating skeleton-based video representations of sign language at the word level. Our approach employs a conditional Generative Adversarial Network (cGAN) with transformer embeddings in both generator and discriminator, augmented with bone-length and joint-angle constraints and a classifier-guided loss to ensure anatomically plausible and semantically consistent gestures. We further introduce a novel loss function to improve human keypoint generation for sign representation. Extensive experiments on three benchmark datasets demonstrate that our method outperforms state-of-the-art approaches according to statistical (MMD) and perceptual (FID) metrics, while qualitative analyses confirm that the generated gestures are temporally smooth, anatomically accurate, and semantically meaningful. These results highlight the effectiveness of our model in advancing word-level sign language synthesis.
Applied Article
Hassan Deldar; Mohammad Mehdi Homayounpour
Abstract
In most of the countries, the legislative process has a long history, which has led to increasing diversity and multiplicity of laws. This has made it difficult to access laws that are valid in both time and place. The focus of this article is on the application of artificial intelligence in the domain ...
Read More
In most of the countries, the legislative process has a long history, which has led to increasing diversity and multiplicity of laws. This has made it difficult to access laws that are valid in both time and place. The focus of this article is on the application of artificial intelligence in the domain of legal statutes to assist in identifying the need for amendments to laws or specific provisions. The general framework of the proposed process consists of two key components.First, the texts of legal clauses or articles are enriched through the generation of enriched data using large language models, which involves producing embedding vectors, thematic classification,and extracting the provisions of each law. Second, a retrieval-augmented text generation (RAG) system is developed with the aid of large language models to determine conflicts or the need for expurgation in the output, utilizing the enriched data, predefined prompts, and the Chain of Thought (CoT) technique.The proposed method was evaluated on two benchmark datasets.On the COLIEE 2025 dataset, our approach outperformed the 2024 winners in legal implication tasks, achieving an F1 score of 0.6521 with minimal prompting. The second evaluation used over 1,000 legal clauses covering abrogation and neutral rules, yielding an impressive F1 score exceeding 73.41%.The findings of the proposed methodology demonstrate that, even with limited expertise in the legal domain, it is possible to identify conflicts and the necessity for refining legal texts to an acceptable degree within a reasonable timeframe for legal experts, leveraging the capabilities of large language models.
Original/Review Paper
Somayyeh Jafarali Jassbi; Sajjad Daliri
Abstract
The rapid growth of the Internet‑of‑Things (IoT) imposes significant challenges on task offloading in fog environments, including service latency, resource constraints, and trust management. Fog computing mitigates these limitations by moving computation and storage closer to end devices. This paper ...
Read More
The rapid growth of the Internet‑of‑Things (IoT) imposes significant challenges on task offloading in fog environments, including service latency, resource constraints, and trust management. Fog computing mitigates these limitations by moving computation and storage closer to end devices. This paper presents BCOFF (Blockchain‑based Computation Offloading Framework for Fog), a secure and efficient framework that jointly optimizes resource allocation and enables verifiable task offloading. In BCOFF, resource allocation is performed using the Grey Wolf Optimization (GWO) algorithm, while blockchain provides a tamper-resistant execution record. Specifically, the blockchain serves three purposes: (i) recording offloading decisions and cryptographic hashes of task results to support post‑execution auditability, (ii) validating the integrity of returned results by matching them with the on‑chain hash reference, and (iii) coordinating consensus among fog nodes through a lightweight Validator‑Selection Proof‑of‑Stake (VNPoS) mechanism. VNPoS is a simplified adaptation of the Nominated Proof‑of‑Stake (NPoS) model that selects validators using stake‑based nomination with variance‑aware stake normalization. By avoiding computationally intensive cryptographic puzzles, VNPoS significantly reduces consensus overhead and is therefore suitable for resource‑constrained fog environments. Experimental evaluation using the iFogSim simulator with workloads of 800–1500 tasks shows that BCOFF reduces execution time by 15–27%, lowers host‑selection latency by 22–25%, and decreases energy consumption by 5–9% compared with existing approaches. These results demonstrate that integrating GWO‑based scheduling with the VNPoS blockchain mechanism provides a more efficient and verifiable fog-offloading framework.
Original/Review Paper
Mohammad M. AlyanNezhadi; Hesamoddin Pourrostami; Mousa Nazari; Farzan Afshari
Abstract
In Iran’s financial market, the authentication of gold coins is majorly required for transparency, reducing fraud, and proper valuation. Differentiating between bank-issued and non-bank-issued coins pose a challenge as their appearance is almost the same. This paper suggests a classification method ...
Read More
In Iran’s financial market, the authentication of gold coins is majorly required for transparency, reducing fraud, and proper valuation. Differentiating between bank-issued and non-bank-issued coins pose a challenge as their appearance is almost the same. This paper suggests a classification method that is based on deep learning and has three main components: extracting area of interest, aligning images through a CNN regressor, and classifying coins through a CNN classifier. The method is tested on a set of 130 coins images (71 coins from banks and 59 coins from non-banks) and is benchmarked against baseline models employing feature extraction and SVMs. The proposed method outperforms the baseline with 99% accuracy. The results prove that the model works effectively in authenticating the coins, which enables safe transactions in the gold market.
Original/Review Paper
Amirhossein Zare Kordkheili; Amirreza Zare Kordkheili; Sekine Asadi Amiri
Abstract
Brain tumor detection is a critical task in medical imaging, requiring accurate and reliable methods. Recent advancements in deep learning have shown great potential in this field. In this article, we present a novel method for brain tumor detection based on a Convolutional Block Attention Module (CBAM) ...
Read More
Brain tumor detection is a critical task in medical imaging, requiring accurate and reliable methods. Recent advancements in deep learning have shown great potential in this field. In this article, we present a novel method for brain tumor detection based on a Convolutional Block Attention Module (CBAM) enhanced attention ensemble of deep learning networks. Initially, image augmentation is applied to increase data diversity. We utilize two deep neural network models, EfficientNet-B1 and ResNet-101, for tumor detection. First, we enhance the performance of these models by integrating the CBAM attention module into their architectures. Then, we ensemble the two networks using a soft voting strategy to achieve higher detection accuracy. The proposed method is evaluated on the three-class Figshare dataset, achieving an accuracy of 99.09% in detecting tumors in MRI images, which outperforms existing methods. This approach leverages the strengths of an ensemble of models, offering a promising solution for improving the accuracy and reliability of brain tumor detection in medical imaging.
Technical Paper
Naga Subrahmanyeswari Nimmakayala; Krishna Prasad M H M
Abstract
Breast cancer detection is critical for early diagnosis and treatment. This paper utilized the BreakHis dataset, comprising 7,907 histopathological images of breast tumors (benign and malignant) captured at varying magnification levels. Initially, a basic CNN was applied, followed by advanced deep learning ...
Read More
Breast cancer detection is critical for early diagnosis and treatment. This paper utilized the BreakHis dataset, comprising 7,907 histopathological images of breast tumors (benign and malignant) captured at varying magnification levels. Initially, a basic CNN was applied, followed by advanced deep learning architectures including ResNet, EfficientNet, Mobilenet, Densenet and VGG19. Among these models, ResNet achieved the highest accuracy of 90.2%. For improving performance, a hybrid combination of hand-crafted features (pHash, HOG, GLCM, Hu Moments, SIFT, ORB and LBP) and transfer learning features (EfficientNet, DenseNet, ResNet, VGG19 MobileNet) was considered. Combined, these features were merged into a single feature vector, and were classified using ML algorithms: Logistic Regression, Naive Bayes, KNN, Decision Tree, Random Forest, Gradient Boosting and XGBoost. XGBoost yielded the highest accuracy of 96.2%. Additionally, deep learning models including Multilayer Perceptron (MLP) and Artificial Neural Networks (ANN) were explored, with ANN slightly outperforming MLP, achieving an accuracy of 98.3% compared to 97.5% for MLP. The results highlight the efficacy of combining traditional and deep learning-based features for improved diagnostic accuracy
Original/Review Paper
Mehdi Alizadeh; Parvin Ahmadi; Masoumeh Azimzadeh
Abstract
Field boundary detection is a critical task in modern agriculture, enabling precision farming, optimized resource management, and efficient crop monitoring. Despite its importance, existing deep learning models often fail to achieve high accuracy in delineating field boundaries due to challenges such ...
Read More
Field boundary detection is a critical task in modern agriculture, enabling precision farming, optimized resource management, and efficient crop monitoring. Despite its importance, existing deep learning models often fail to achieve high accuracy in delineating field boundaries due to challenges such as complex landscapes, varying resolutions, and noise in remote sensing images. To overcome these limitations, we propose HURA-Net, an advanced deep learning framework that intelligently integrates UNet++, ResUNet, and an attention mechanism into a unified architecture. By hybridizing these models, HURA-Net effectively combines their strengths—such as multi-scale feature extraction (UNet++), residual learning (ResUNet), and focus on salient regions (attention mechanism)—while minimizing their individual weaknesses. To further enhance performance, we introduce a refined loss function that not only improves segmentation precision but also addresses the class imbalance problem, which is common in boundary detection tasks. Extensive experiments on a diverse dataset of high-resolution satellite images from different regions of Iran demonstrate that HURA-Net significantly outperforms existing state-of-the-art models. Specifically, it achieves a recall of 45.85% (a 15.59% improvement over ResUNet) and an F1-score of 42.62% (7.27% higher than ResUNet), setting a new benchmark for accuracy. Moreover, our study highlights the critical role of strategic data augmentation in boosting model generalization, particularly in handling variations in lighting, crop types, and field shapes. The success of HURA-Net underscores the importance of innovative architecture design, optimized loss functions, and robust training strategies in advancing remote sensing image segmentation.
Original/Review Paper
Samira Mavaddati
Abstract
The rapid growth of intelligent surveillance systems has increased the demand for accurate and efficient criminal activity recognition methods capable of operating in real-world environments. Although conventional deep learning and object detection frameworks have demonstrated promising performance, ...
Read More
The rapid growth of intelligent surveillance systems has increased the demand for accurate and efficient criminal activity recognition methods capable of operating in real-world environments. Although conventional deep learning and object detection frameworks have demonstrated promising performance, they often struggle to capture long-range contextual dependencies and complex interactions present in surveillance scenes. To address these limitations, this study proposes a hybrid deep learning framework that combines the real-time detection capability of YOLOv10 with the global contextual modeling power of Vision Transformers (ViT). An attention-guided feature fusion mechanism is introduced to effectively integrate local spatial representations extracted by YOLOv10 with global semantic features generated by the transformer architecture. The proposed framework is evaluated on the UCF-Crime dataset, which consists of fourteen categories of normal and criminal activities, including burglary, robbery, assault, vandalism, shoplifting, and abuse. Surveillance videos are converted into image sequences and analyzed under two experimental scenarios: (I) a standalone YOLOv10 model and (II) the proposed Attention-Guided YOLOv10-ViT framework. Performance is assessed using accuracy, precision, recall, and F1-score metrics. Experimental results show that the standalone YOLOv10 model achieves an overall classification accuracy of 88.07%, outperforming the previously reported YOLOv8 baseline. More importantly, the proposed hybrid framework attains an accuracy of 93.45%, exceeding both YOLOv10 and earlier YOLOv8-ViT architectures. The improvement is particularly evident in challenging scenarios involving occlusion, illumination changes, cluttered backgrounds, and crowded environments. The results demonstrate that integrating YOLOv10, Transformers, and attention-guided feature fusion provides a scalable, robust, and real-time solution for intelligent surveillance and public monitoring applications.
Original/Review Paper
Fatemeh Naghavi; Kourosh Kiani
Abstract
Monitoring the daily activities of elderly individuals plays a crucial role in accident prevention, health assessment, and improving quality of life. In this paper, we propose a lightweight and efficient convolutional neural network architecture for human activity recognition based on skeletal data. ...
Read More
Monitoring the daily activities of elderly individuals plays a crucial role in accident prevention, health assessment, and improving quality of life. In this paper, we propose a lightweight and efficient convolutional neural network architecture for human activity recognition based on skeletal data. Unlike conventional approaches that rely solely on absolute joint coordinates, the proposed method incorporates short- and long-term frame differences as well as spatial variations across joints to construct complementary views, thereby providing a richer spatiotemporal representation. The architecture consists of multiple convolutional blocks with residual connections, followed by global average pooling and a fully connected layer for final classification. Experimental evaluations conducted on two benchmark datasets, NTU RGB+D and ETRI-Activity3D, demonstrate that while the proposed model may achieve slightly lower accuracy compared to some state-of-the-art methods, it offers high inference speed and low computational complexity. These characteristics make it particularly suitable for real-time applications and deployment on resource-constrained devices, especially in elderly home-care environments.
Methodologies
Atefeh Niroomand; Seyyed Hamid Ghafouri; Amid Khatibi Bardsiri
Abstract
This study addresses the challenges of managing dynamic and heterogeneous Internet of Things (IoT) data by proposing a time-aware recommender system that integrates a dynamic semantic ontology with clustering techniques and a hybrid collaborative filtering framework. The proposed model continuously updates ...
Read More
This study addresses the challenges of managing dynamic and heterogeneous Internet of Things (IoT) data by proposing a time-aware recommender system that integrates a dynamic semantic ontology with clustering techniques and a hybrid collaborative filtering framework. The proposed model continuously updates the ontology based on user interactions and incorporates temporal information into both knowledge representation and clustering processes, enabling adaptive and real-time modeling of evolving user behaviors.The dataset consists of approximately 500 users and 15,000 time-stamped interaction records collected over four months, including demographic attributes (age and gender), IoT device usage patterns, and temporal features such as timestamp and time of day.The recommendation framework combines ontology-enhanced user-based collaborative filtering with dynamic K-means clustering, leveraging both semantic relationships and behavioral similarities to improve recommendation quality. Experimental evaluation is conducted using Precision, Recall, F1-score, Accuracy, MAE, and RMSE metrics. The model achieves improvements ranging from approximately 2% to 52%, with respect to state-of-the-art non-temporal methods and traditional collaborative filtering techniques, respectively.Furthermore, computational complexity analysis indicates that the additional processing cost introduced by dynamic ontology updates and temporal modeling remains manageable, preserving the practical applicability of the proposed framework in resource-constrained IoT environments.
Research Note
Mohammad Heydari
Abstract
Deep learning–based super-resolution has become an important tool for enhancing brain magnetic resonance imaging (MRI), particularly when acquisition constraints limit spatial resolution. Lightweight autoencoder architectures have recently been proposed to achieve computational efficiency while ...
Read More
Deep learning–based super-resolution has become an important tool for enhancing brain magnetic resonance imaging (MRI), particularly when acquisition constraints limit spatial resolution. Lightweight autoencoder architectures have recently been proposed to achieve computational efficiency while maintaining reconstruction quality. However, certain architectural choices adopted in these models—most notably the incorporation of encoder–decoder skip connections—raise methodological concerns regarding the preservation of the information bottleneck principle that defines autoencoder-based learning. This note critically examines whether such designs genuinely rely on latent representation learning or instead introduce shortcut pathways that weaken the inferential nature of super-resolution reconstruction.
Original/Review Paper
Omid Khalaf Beigi; Seyed Alireza Bashiri Mosavi
Abstract
A brain tumor is one of the most serious and life-threatening brain diseases that can profoundly affect an individual’s life. Accordingly, the present study addresses the challenge of refining brain tumor segmentation based on Magnetic Resonance Imaging (MRI) data and deep reinforcement learning. ...
Read More
A brain tumor is one of the most serious and life-threatening brain diseases that can profoundly affect an individual’s life. Accordingly, the present study addresses the challenge of refining brain tumor segmentation based on Magnetic Resonance Imaging (MRI) data and deep reinforcement learning. Although supervised learning–based approaches have shown satisfactory performance in tumor segmentation and localization, they often suffer from high uncertainty errors along tumor boundaries. In this research, a learning framework combining a supervised model with deep reinforcement learning—referred to as DURL-Net—is proposed for segmentation and refinement purposes. Specifically, the framework first employs a U-Net architecture to generate an initial segmentation mask. This initial output and the corresponding MRI are then partitioned into localized patches, which are sequentially processed by a Deep Q-Network (DQN) agent. The DQN agent interacts with the environment by selecting optimal morphological operations (such as dilation and erosion) to refine tumor boundaries and correct uncertainties patch by patch. The dataset used in this study comprises 3,064 T1-Weighted Contrast-Enhanced MRI images, employed for both segmentation and tumor-type classification tasks. Experimental results demonstrate that DURL-Net achieved a Dice Similarity Coefficient (DSC) of 86.73%, a Jaccard Index (IoU) of 78.68%, a Kappa coefficient (Kap) of 85.21%, a Sensitivity of 87.68%, and a Specificity of 96.06%.
Applied Article
Mohammad Aalishahi; Mohammad Hadi Bokaei; Abolfazl Nadi
Abstract
Given the importance of Named Entity Recognition (NER), numerous studies have been conducted in this field. However, most research has focused on languages such as English, French, and Arabic. In contrast, studies on Persian remain limited, despite Persian being one of the most widely spoken languages ...
Read More
Given the importance of Named Entity Recognition (NER), numerous studies have been conducted in this field. However, most research has focused on languages such as English, French, and Arabic. In contrast, studies on Persian remain limited, despite Persian being one of the most widely spoken languages in West Asia, necessitating the development of NER methods for it. In this study, using Active Learning, a corpus of 1,351 advertisements from the Official Gazette was annotated. The GEMMA2b model was then fine-tuned on this data, achieving approximately 95% accuracy. This model was employed to extract around 13 types of named entities and their relationships within the advertisement texts. The primary advantage of this method is the model’s high accuracy compared to other approaches. Additionally, the use of Persian data—which, unlike languages such as English or Arabic, has fewer resources—is another notable feature of this research.
Conceptual Paper
farzane maghsoudi; Mohammad Javad fadaeiEslam; Farzin Yaghmaee
Abstract
Image-to-image translation is a highly challenging task, as it requires an accurate understanding of image details and their consistent transformation across domains. Notably, GANs have achieved remarkable success in this field. In essence, convolutional layers are the primary building blocks of these ...
Read More
Image-to-image translation is a highly challenging task, as it requires an accurate understanding of image details and their consistent transformation across domains. Notably, GANs have achieved remarkable success in this field. In essence, convolutional layers are the primary building blocks of these architectures. However, the limited receptive field in shallow layers makes it difficult to capture long-range spatial dependencies and non-local context. In this paper, the HiSGAN architecture is proposed to address this limitation. It combines deep representations with traditional techniques, such as SVD and Fast Fourier Convolution (FFC), to effectively extract style-related information and establish a global receptive field. Furthermore, we introduce the HiS-Transformer block with an involution operator in the bottleneck of the generator. This proposed block utilizes hybrid-scale self-attention to adaptively preserve the global receptive field and fine-grained information in salient regions while maintaining low computational cost. HiSGAN employs a new loss function based on gradient contrastive learning to improve cross-domain feature alignment. Quantitative and qualitative results on four public datasets demonstrate the superiority of the proposed approach over state-of-the-art methods. Importantly, these performance gains are achieved while reducing the parameter count and accelerating training. The code is available at https://github.com/OliverRensu/SG-Former
Other
Mohammad Mahdi Estesnaei
Abstract
Spectrum-based fault localization (SBFL) is a widely used technique that utilizes coverage data and test outcomes to calculate a suspiciousness score for each program statement. The fundamental hypothesis of SBFL is that a statement covered by more failed test cases and fewer passed test cases is more ...
Read More
Spectrum-based fault localization (SBFL) is a widely used technique that utilizes coverage data and test outcomes to calculate a suspiciousness score for each program statement. The fundamental hypothesis of SBFL is that a statement covered by more failed test cases and fewer passed test cases is more likely to be faulty. However, the effectiveness of SBFL is hindered by coincidental correctness, which occurs when a fault is executed but no failure is detected. Additionally, traditional SBFL methods assign equal weight to all failed tests, despite some failed tests containing more valuable information. This study aims to enhance SBFL performance by employing a fuzzy expert system to address these challenges. Thirteen open-source subject programs were used to evaluate the efficiency of the proposed FSBFL method. Experimental results, assessed using four key metrics, demonstrate that FSBFL outperforms popular spectrum-based fault localization techniques.
Original/Review Paper
Niloofar Ranjbar; Hamed Baghbani
Abstract
Retrieval-augmented generation (RAG) is commonly evaluated on clean inputs that underrepresent realistic multilingual variation. We present an English-Persian movie-domain robustness benchmark built from a corpus of 31,564 records, 120 clean queries, and 720 aligned perturbations. The benchmark covers ...
Read More
Retrieval-augmented generation (RAG) is commonly evaluated on clean inputs that underrepresent realistic multilingual variation. We present an English-Persian movie-domain robustness benchmark built from a corpus of 31,564 records, 120 clean queries, and 720 aligned perturbations. The benchmark covers six deterministic query types and 14 operational perturbation labels grouped into four families. We compare BM25, multilingual dense retrieval, character n-gram TF-IDF, and hybrid retrieval, and evaluate top-1 deterministic answer extraction against a field-specific top-5 RAG system using Qwen2-7B-Instruct. Hybrid retrieval achieves 81.50 MRR@10 on clean queries and 67.76 under perturbation; field-specific RAG reaches 84.17% and 72.08% accuracy, respectively. Clustered paired-bootstrap 95% confidence intervals exclude zero for all principal system differences. English-title noise is the most damaging family, whereas query-form and punctuation variation is comparatively well tolerated. A 43-case consistency audit verifies implementation of the rule-based failure categories, and full-output analysis shows that retrieval-coverage errors dominate the difficult English-title family. These results support component-level evaluation of multilingual RAG robustness.
Original/Review Paper
alireza Omidi nasab; Sajad Bastami; Rojiar Pir Mohammadiani; Mohammad Bagher Dowlatshahi; Seyedeh Zahra Mousavi
Abstract
Deep Neural Networks (DNNs) are increasingly deployed in safety-critical domains such as autonomous driving, healthcare, finance, and natural language processing, yet they remain vulnerable to adversarial attacks—subtle manipulations that can cause confident misclassifications or misleading predictions. ...
Read More
Deep Neural Networks (DNNs) are increasingly deployed in safety-critical domains such as autonomous driving, healthcare, finance, and natural language processing, yet they remain vulnerable to adversarial attacks—subtle manipulations that can cause confident misclassifications or misleading predictions. This fragility poses a major barrier to building secure and trustworthy AI systems. Conventional defenses, including adversarial training and heuristic detection, often struggle to balance robustness, adaptability, and computational cost. To overcome these limitations, we propose a hybrid adaptive defense framework that unifies Ant Colony Optimization (ACO) with Reinforcement Learning (RL). ACO efficiently explores the high-dimensional space of defense hyperparameters to find globally optimal configurations, while RL enables dynamic, context-aware adaptation of defense strategies in real time. The proposed ACO-RL framework was rigorously evaluated across six diverse benchmark datasets spanning multiple data modalities: MNIST and CIFAR-10 (vision), IMDB and AG News (text), and Cora and Reddit-Binary (graph). Experimental results show that ACO-RL consistently enhances robustness against a wide spectrum of adversarial attacks, outperforming several state-of-the-art baselines. These findings highlight a promising pathway toward developing resilient, cross-domain AI systems capable of defending against evolving adversarial threats.
Original/Review Paper
omid hajipoor; Ahmad Nickabadi; Mohammad Mehdi Homayounpour
Abstract
Conditional text generation is crucial in natural language processing but often struggles with the high computational costs of Large Language Models (LLMs) and training instability in Generative Adversarial Networks (GANs). In this paper, we introduce the Guidance-Transformer Generative Adversarial Network ...
Read More
Conditional text generation is crucial in natural language processing but often struggles with the high computational costs of Large Language Models (LLMs) and training instability in Generative Adversarial Networks (GANs). In this paper, we introduce the Guidance-Transformer Generative Adversarial Network (GTGAN), a framework that generates text by working within a smooth, continuous hidden (latent) space rather than outputting discrete words directly, bypassing common optimization bottleneck errors. To enforce precise control, GTGAN utilizes a dual-guidance classifier mechanism to organize this latent space and direct the generation process. Additionally, we employ a global discriminator (to maintain overall sentence coherence) and a local discriminator (to verify small word groups), which together prevent repetitive text (mode collapse) and training errors (exposure bias). Tested on Yelp and Amazon reviews, GTGAN improves text quality (BLEU-5) by up to 18.9% and category retention accuracy by 8.78% over preceding GANs. Notably, GTGAN achieves these results while using under 50M parameters (less than 1% of the size of standard 7B parameter LLMs), demonstrating a highly efficient, controllable, and lightweight solution for conditional text generation.
Original/Review Paper
AGHIL Kashir Taghartapeh; Nader Javadifar; Ali Harimi; Seyed Mehdi BagheriMofidi; Aziz Kalteh
Abstract
This study investigates the effectiveness of integrating nonlinear dynamical representations derived from reconstructed phase space (RPS) analysis with deep convolutional neural networks for phonocardiogram classification. It evaluates how the nonlinear dynamic information present in cardiac signals ...
Read More
This study investigates the effectiveness of integrating nonlinear dynamical representations derived from reconstructed phase space (RPS) analysis with deep convolutional neural networks for phonocardiogram classification. It evaluates how the nonlinear dynamic information present in cardiac signals can be captured through the proposed unified framework that embeds RPS within a deep learning architecture. Utilizing the ResNet-34 model, the method efficiently extracts critical spatial and temporal patterns from these enhanced tensors, offering a novel integration of nonlinear system dynamics with deep feature learning for improved diagnostic performance. Performance evaluation conducted on the PhysioNet Challenge dataset demonstrates the method's capability to achieve high accuracy, precision, recall, and F1 scores consistently across different tensor dimensions and time lag parameters. Best results were obtained using tensors with dimension 4 and time lags of 3 and 5, underlining the model's robustness and flexibility in accommodating variability inherent to PCG signals. With an F1 score of 93.38%, the proposed method performs competitively within the upper range of current state-of-the-art techniques.
Research Note
vahidreza afshin; Saiedeh Kabirirad; Seyed Hamid Zahiri
Abstract
Accurate and timely detection of heart disease is a fundamental challenge in clinical diagnostics, exacerbated by the inherent complexities of real-world healthcare datasets, such as class imbalance, high dimensionality and the presence of outliers. To address these issues, we propose a dual-path learning ...
Read More
Accurate and timely detection of heart disease is a fundamental challenge in clinical diagnostics, exacerbated by the inherent complexities of real-world healthcare datasets, such as class imbalance, high dimensionality and the presence of outliers. To address these issues, we propose a dual-path learning framework specifically designed for real-time heart disease detection within IoT–Fog–Cloud environments. The framework integrates advanced preprocessing techniques, including Conditional Generative Adversarial Networks (CGANs) for data balancing and a hybrid feature selection pipeline, thereby addressing data imbalance and dimensionality reduction. It also proposes a dual-path diagnostic system that uses a lightweight XGBoost classifier in the fog layer optimized for low latency and fast local inference, and an improved TabTransformer-based model in the cloud layer, which is selectively used for ambiguous samples. Experimental evaluation on heart disease datasets demonstrates the superior performance of the framework, achieving 99.66% accuracy, F1 score of 99.75%, and AUC of 99.96%. Additionally, it improves overall inference latencies, mostly below one millisecond in the fog layer and low latency in cloud layer. This paper provides a scalable, interpretable, and privacy-preserving medical diagnosis framework in distributed healthcare systems.
Original/Review Paper
Hamed Modanloujouybari; Yasser Baleghi
Abstract
Accurate estimation of velocity fields from Particle Image Velocimetry (PIV) data is essential for fluid-flow analysis and modeling. PIV relies on Image Processing techniques such as cross-correlation and optical flow to estimate the magnitude and direction of fluid motion; however, traditional approaches ...
Read More
Accurate estimation of velocity fields from Particle Image Velocimetry (PIV) data is essential for fluid-flow analysis and modeling. PIV relies on Image Processing techniques such as cross-correlation and optical flow to estimate the magnitude and direction of fluid motion; however, traditional approaches often struggle in the presence of noise, sharp velocity gradients, and complex flow structures. Recent deep-learning-based methods have shown promising performance, but purely data-driven models may generate physically inconsistent predictions because they do not explicitly incorporate the governing laws of fluid dynamics. In this study, a physics-informed framework based on Recurrent All-Pairs Field Transforms (RAFT), a state-of-the-art deep neural network for dense optical flow estimation, is proposed for velocity-field reconstruction from PIV image sequences. The proposed approach introduces a novel loss function that combines a Charbonnier data term with divergence-free, vorticity-based, and edge-aware smoothness constraints to improve the physical consistency and robustness of the estimated flow fields. The method was evaluated on five benchmark PIV datasets and consistently outperformed the original RAFT model. The average endpoint error (AEE) was reduced from 16.07 to 14.19, corresponding to an improvement of approximately 11.7%, with the largest gains observed in turbulent and complex flow conditions. These results demonstrate that incorporating fluid-dynamics knowledge into deep-learning-based optical flow estimation improves the accuracy, stability, and generalization capability of PIV velocity-field reconstruction.
Original/Review Paper
Sajad Bastami; Mohammad Bagher Dowlatshahi; Rojiar Pir Mohammadiani; Seyedeh Zahra Mousavi
Abstract
Multi-agent reinforcement learning (MARL) is a key paradigm for coordination in robotics, autonomous systems, and distributed control. However, existing MARL methods face fundamental limitations in scalability, adaptability to dynamic environments, and stability under evolving interactions. To address ...
Read More
Multi-agent reinforcement learning (MARL) is a key paradigm for coordination in robotics, autonomous systems, and distributed control. However, existing MARL methods face fundamental limitations in scalability, adaptability to dynamic environments, and stability under evolving interactions. To address these challenges, we propose Adaptive Graph-Transformer Reinforcement Learning (AGTRL), a framework integrating graph-based relational modelling with transformer attention for adaptive coordination in large-scale multi-agent systems. AGTRL unifies graph-based perception and attention-based coordination in an end-to-end pipeline, encoding role information and adaptively weighting interactions by context. This combination, missing in prior MARL methods, bridges scalability and robustness in dynamic environments. AGTRL constructs a dynamic graph of agent relationships and uses multi-head self-attention to prioritize relevant interactions in real-time, ensuring robust performance under perturbations. We evaluate robustness under communication dropout (up to 40% link removal) and dynamic edge removal, measuring performance via episode reward and win rate. The framework incorporates an adaptive stability-performance trade-off mechanism that maintains learning efficacy in the presence of communication constraints and environmental uncertainty. We introduce a graph-enhanced policy architecture that jointly optimizes individual agent policies and inter-agent coordination through attention-weighted message passing. Comprehensive evaluations on benchmark environments—including StarCraft II micromanagement scenarios, cooperative navigation (Spread), and adversarial tasks (Predator-Prey)—demonstrate that AGTRL achieves superior sample efficiency, scalability, and robustness compared to state-of-the-art MARL baselines. Experimental results show AGTRL improves convergence speed by 32% on average and maintains stable performance with up to 40% communication dropout, establishing its viability for real-world deployment in dynamic multi-agent domains.
Original/Review Paper
Yashar Chehardahcherikigheisari; Hadi Grailu
Abstract
Human emotion recognition based on electroencephalogram signals remains a significant challenge in computational neuroscience and artificial intelligence. Convolutional neural networks have been widely adopted to address this challenge due to their strong capabilities in feature extraction and representation ...
Read More
Human emotion recognition based on electroencephalogram signals remains a significant challenge in computational neuroscience and artificial intelligence. Convolutional neural networks have been widely adopted to address this challenge due to their strong capabilities in feature extraction and representation learning. Empirical tuning of hyperparameters, however, is often inefficient and prone to suboptimal solutions. To overcome this limitation, the Grey Wolf Optimizer and Whale Optimization Algorithm—two swarm intelligence–based metaheuristic methods—were employed to refine the architecture and learning hyperparameters of a baseline convolutional neural network. Performance was evaluated on the SEED EEG dataset for emotion recognition. Experimental results showed that the optimized model achieved an accuracy of 86% compared to 83% for the baseline, with the confusion matrix confirming reduced misclassification errors and improved recognition of emotional states. Beyond accuracy, the optimization process also lowered the number of trainable parameters and computational overhead, thereby enhancing efficiency. These findings highlight that swarm intelligence–based methods provide superior exploration and exploitation capabilities, enabling systematic hyperparameter tuning and delivering a balanced engineering approach that combines predictive performance with resource efficiency.
Research Note
Mostafa Ghazizadeh-Ahsaee; Afsaneh Shamsaddini-Farsangi
Abstract
Clustering datasets with varying densities is challenging because classical Density Peak Clustering (DPC) may fail to identify valid centers in sparse regions and may incorrectly assign sparse samples to dense clusters. To address this issue, this paper proposes a Density Peak Clustering algorithm based ...
Read More
Clustering datasets with varying densities is challenging because classical Density Peak Clustering (DPC) may fail to identify valid centers in sparse regions and may incorrectly assign sparse samples to dense clusters. To address this issue, this paper proposes a Density Peak Clustering algorithm based on Fuzzy Shared Nearest Neighbors (DPC-FSNN). The proposed method combines k-nearest neighbors and fuzzy neighborhood information to construct a fuzzy nearest neighbor kernel for adaptive local density estimation. It also incorporates shared nearest neighbor information into the local distance computation and sample assignment process to improve center detection and assignment reliability. Experiments on varying-density datasets (e.g., Jain, Compound, and LineBlobs), complex-shape datasets, and real-world datasets are evaluated using AMI, ARI, and FMI. The results show that DPC-FSNN achieves competitive and often superior clustering performance compared with IDPC-FA, FKNN-DPC, DPC-FWSN, DPCSA, FNDPC, and the classical DPC algorithm. A computational complexity analysis is also provided, showing that the proposed method has overall complexity O(n^2) in the standard implementation.
Original/Review Paper
Matin Gholami; Seyyed Ali Zendehbad; Jamal Ghasemi
Abstract
Deep neural networks are data-hungry and prone to overfitting, especially with limited training data and fine-grained visual variations. While region-mixing augmentations like CutMix serve as effective regularizers, their purely stochastic nature often creates a semantic gap—critical discriminative ...
Read More
Deep neural networks are data-hungry and prone to overfitting, especially with limited training data and fine-grained visual variations. While region-mixing augmentations like CutMix serve as effective regularizers, their purely stochastic nature often creates a semantic gap—critical discriminative features are inadvertently masked or replaced, yielding suboptimal training signals. To address this limitation, the authors propose a novel family of adaptive augmentation strategies guided by Grad-CAM attention maps, injecting semantic awareness into the mixing procedure. The framework comprises three variants: Feature-Based CutMix and Saliency-Guided CutMix, which target salient regions for replacement (single peak and multiple high-saliency regions via fixed thresholding, respectively); and Region-Adaptive CutMix (RACutMix), which preserves salient regions by generating non-rectangular, saliency-informed masks using an adaptive percentile-based thresholding mechanism. RACutMix strategically augments only background or low-saliency regions with donor content, producing semantically coherent training examples that maintain visual and label consistency. Systematic evaluation on the Oxford-IIIT Pet benchmark with ResNet-50 compares five training paradigms under identical protocols with repeated trials and significance testing. RACutMix achieves 89.76% test accuracy, a substantial 10.23 percentage point improvement over standard CutMix (79.53%). It also demonstrates faster convergence, more stable loss dynamics, and consistent gains across precision, recall, and F1-score. Training overhead increases by approximately 1.5× time and 0.5 GB peak memory—a cost the authors quantify fully. This work bridges explainable AI and training regularization, demonstrating that internal model attention can be effectively recycled to guide learning toward more robust and discriminative feature representations, validating interpretability-guided augmentation as a powerful paradigm for fine-grained visual recognition tasks.
Technical Paper
Mahboobeh Riahi-Madvar; zahra Moterassed
Abstract
Anomaly detection in healthcare datasets is vital for identifying unusual and potentially critical patterns that can influence clinical decisions. However, because many medical datasets are small and have high dimensionality, this task is still a challenge. While deep neural networks have achieved significant ...
Read More
Anomaly detection in healthcare datasets is vital for identifying unusual and potentially critical patterns that can influence clinical decisions. However, because many medical datasets are small and have high dimensionality, this task is still a challenge. While deep neural networks have achieved significant success in large-scale data environments, their performance often diminishes when applied to small datasets. Recently, ensemble approaches have gained considerable attention because of their capacity to combine the results of multiple detectors for more robust anomaly detection. In this study, we present an Anomaly Detection approach with Bagging-Principal Component Analysis (ADBP) specifically designed for small, high-dimensional healthcare datasets. Our approach integrates bagging sampling with Principal Component Analysis to create multiple diverse feature spaces, effectively addressing the challenges of limited data and high dimensionality. Specifically, several subsets of the training data are randomly sampled to learn separate Principal Component Analysis subspaces. Within each subspace, the K Nearest Neighbors algorithm is employed to detect anomalies. Final anomaly detection is obtained through a voting mechanism that aggregates the predictions from all detectors. Our ADBP method was evaluated on eight public UCI datasets-Arrhythmia, Parkinson, Annthyroid, Cardiotocography, Hepatitis, Lymphography, Heart Disease, and Pima. Our experimental results show that ADBP algorithm outperforms several state-of-the-art methods, especially in scenarios involving small datasets. The results confirm that integrating bagging, Principal Component Analysis, and ensemble learning techniques significantly improve detection performance in high-dimensional, small datasets. These findings highlight the potential of the ADBP approach as a reliable anomaly detection in healthcare applications.
Original/Review Paper
Zahrasadat Sajjadi; Soheil Hamzebeigi; Mohsen Soryani
Abstract
As the global population ages, reliable methods for assessing brain health and age-related changes are increasingly important. Brain age is a promising biomarker of brain health, and machine-learning methods have enabled its estimation from neuroimaging data. However, effective training strategies are ...
Read More
As the global population ages, reliable methods for assessing brain health and age-related changes are increasingly important. Brain age is a promising biomarker of brain health, and machine-learning methods have enabled its estimation from neuroimaging data. However, effective training strategies are required for accurate brain age estimation. This study proposes a two-dimensional convolutional neural network (2D CNN)-based multi-task framework for estimating brain age from magnetic resonance imaging (MRI) scans in the ADNI dataset. The framework uses VGG-16 and U-Net encoder backbones with separate heads for age regression and image-rotation classification. We evaluated the effects of several pretraining strategies, including self-supervised pretraining using the DINO framework and supervised pretraining through brain-tumor segmentation. The best-performing configuration, consisting of a DINO-pretrained VGG-16 backbone and a multi-task prediction head, achieved a mean absolute error (MAE) of 3.27 years, which is competitive with previously reported methods. The results indicate that combining transfer learning with multi-task learning improved performance relative to the corresponding single-task models, suggesting that this combination supports the learning of richer and more generalizable feature representations.
Original/Review Paper
Zainab Hasan; Esmaeel Tahanian
Abstract
Today, telecommunications fraud has emerged as a major challenge for operators, resulting in billions of dollars in financial losses annually. the presence of substantial noise and severe class imbalance between legitimate and fraudulent data complicates the identification of fraud patterns within massive ...
Read More
Today, telecommunications fraud has emerged as a major challenge for operators, resulting in billions of dollars in financial losses annually. the presence of substantial noise and severe class imbalance between legitimate and fraudulent data complicates the identification of fraud patterns within massive volumes of Call Detail Records (CDRs). This paper proposes a hybrid ensemble model, termed Hybrid AdaBoost-RF, for telecommunication fraud detection. In this model, Random Forest is employed as the base learner within the AdaBoost framework to enhance the model's robustness against noise. Furthermore, the SMOTE technique is utilized to address the class imbalance problem. Additionally, we applied a decision threshold tuned on the training predictions to improve the model's sensitivity in detecting fraudulent behavior. Experimental results demonstrate that the proposed model outperforms existing methods in recent research, achieving a Recall of 0.87 and an F1-Score of 0.86 on the test partition of the evaluated CDR dataset using the adopted experimental protocol achieving a Recall of 0.87 and an F1-Score of 0.86. Moreover, the Area Under the Curve (AUC) for ROC and PR metrics reach 0.9777 and 0.8733, respectively, validating the high efficiency of the proposed model.
Applied Article
Kiana Rezaei Jafari; Omid Mahdi Ebadati E.; Hamza Khastar
Abstract
The rapid growth of e-commerce has led to an increasing volume of Persian user reviews containing valuable opinions about products and services. Sentiment analysis enables automatic extraction of sentiment polarity from such data; however, Persian remains underexplored, especially in real-world e-commerce ...
Read More
The rapid growth of e-commerce has led to an increasing volume of Persian user reviews containing valuable opinions about products and services. Sentiment analysis enables automatic extraction of sentiment polarity from such data; however, Persian remains underexplored, especially in real-world e-commerce contexts. In addition, class imbalance in sentiment datasets poses a major challenge, often leading to biased models that underperform on minority classes. In this study, we investigate sentiment analysis on Persian e-commerce reviews by comparing classical machine learning models, including Logistic Regression and SVM, with transformer-based models, namely ParsBERT and ParsRoBERTa. To mitigate the impact of class imbalance, we evaluate Focal Loss against the standard Cross-Entropy Loss. Furthermore, we employ Integrated Gradients within an Explainable Artificial Intelligence (XAI) framework to improve model interpretability and analyze feature contributions. Experimental results on the Digikala dataset demonstrate that ParsBERT trained with Focal Loss achieves the best performance, reaching a Balanced Accuracy of 88.64% and an AUC-ROC of 95.91%. The findings highlight the effectiveness of combining imbalance-aware loss functions with transformer-based architectures for improving minority class detection in Persian sentiment analysis.
Original/Review Paper
Fateme Namazi; Mehdi Ezoji; Ebadat Ghanbari Parmehr
Abstract
Accurate rice mapping is crucial for food security, water management, and long-term agricultural planning. This study proposes a phenology-informed LSTM framework that integrates multi-source Sentinel-1 and Sentinel-2 time-series data for robust rice mapping. Sentinel-2 optical images and Sentinel-1 ...
Read More
Accurate rice mapping is crucial for food security, water management, and long-term agricultural planning. This study proposes a phenology-informed LSTM framework that integrates multi-source Sentinel-1 and Sentinel-2 time-series data for robust rice mapping. Sentinel-2 optical images and Sentinel-1 SAR images were preprocessed using cloud-pixel removal, 10-day temporal compositing, and interpolation of missing observations to generate complete and temporally consistent 18-step time series over a 180-day observation period. Three indices, including Sentinel-2-derived NDVI and LSWI and Sentinel-1-derived CR, were extracted to represent vegetation greenness, surface/vegetation moisture, and radar backscatter dynamics. To incorporate rice phenological prior knowledge, reference rice phenology curves were constructed for these indices, and a sliding-window correlation analysis was applied between each pixel-level index trajectory and the corresponding reference curve. This process generated three correlation-based temporal similarity sequences, C_NDVI, C_LSWI, and C_CR, which describe how closely each pixel follows the characteristic rice growth pattern. The raw indices, correlation-based features, and their combined representation were then evaluated as sequential inputs to an LSTM classifier. The results showed that the correlation-based feature set achieved the best performance, with an overall accuracy of 99.81%, an error rate of 0.19%, and an F1-score of 0.9950. Compared with raw time-series indices and the combined feature set, the correlation-based representation improved the discrimination of rice paddies from spectrally similar classes such as water bodies and weeds. The results demonstrate that integrating phenology-informed correlation features with LSTM-based temporal learning provides an effective framework for accurate rice mapping in fragmented and cloud-prone agricultural landscapes.
Original/Review Paper
Mahdi Ahmadlou; Abolghasem Daeichian; Ali Reihanian
Abstract
The expansion of e-commerce has changed customer purchasing habits, moving them from brick-and-mortar stores to online venues. In this shift, some fundamental customer behaviors had to change because online shoppers cannot physically feel the products and rely heavily on customer reviews for evaluations. ...
Read More
The expansion of e-commerce has changed customer purchasing habits, moving them from brick-and-mortar stores to online venues. In this shift, some fundamental customer behaviors had to change because online shoppers cannot physically feel the products and rely heavily on customer reviews for evaluations. However, the lack of structured textual data poses challenges in sifting through numerous, diverse, and sometimes contradictory comments to make an informed purchasing decision. This study proposes a multi-step system for autonomously and intelligently analyzing customer comments to organize comment sections, utilizing general-purpose large language models. First, the proposed system automatically separates comments that discuss online shop services from those specifically related to products by tagging them. Then, it extracts the various product aspects discussed across all comments. Finally, these comments are categorized based on the extracted aspects. Additionally, a new labeled non-English dataset has been created as a benchmark dataset featuring tagged online-shop-related comments. The experimental results showed that the best performing model was Qwen 2.5, achieving an accuracy of 91.7 %.
)
