Original/Review Paper
H.6.2.4. Neural nets
Ali Asadi; Morteza Noferesti
Abstract
Hydrogen combustion has emerged as a pivotal technology for decarbonizing the energy sector, offering a clean and sustainable alternative to fossil fuels. This study investigates hydrogen combustion dynamics in a perfectly stirred reactor (PSR) under steady-state, non-premixed conditions. It is employing ...
Read More
Hydrogen combustion has emerged as a pivotal technology for decarbonizing the energy sector, offering a clean and sustainable alternative to fossil fuels. This study investigates hydrogen combustion dynamics in a perfectly stirred reactor (PSR) under steady-state, non-premixed conditions. It is employing CHEMKIN-based simulations to analyze the effects of nitrogen dilution, operating pressure, and equivalence ratio on flame temperature and NOx emissions. The parametric results reveal that nitrogen dilution reduces flame temperature by up to 28% and suppresses NOx emissions by 15–40%, while elevated pressure promotes higher flame temperatures and increased NOx formation. Peak temperature and NOx concentrations are observed under stoichiometric conditions (φ = 1.0), with both quantities decreasing under lean and rich mixture conditions. To enable rapid and accurate prediction of these combustion characteristics, three machine learning models were developed and benchmarked on the CHEMKIN-generated dataset: Gaussian Process Regression, Multilayer Perceptron, and Deep Neural Network. GPR demonstrated the best overall predictive performance, achieving the lowest MAE for temperature prediction (MAE = 1.77) and major species concentrations. Although the DNN produced competitive results (MAE = 58.28), it demanded approximately five times more computational resources than GPR, without a proportional gain in predictive accuracy. All three models, maintained prediction errors below 20% across the investigated parameter space, confirming their viability as efficient tools for hydrogen combustion optimization. These findings demonstrate that physics-informed machine learning models, when combined with high-fidelity combustion simulations, offer a powerful and computationally efficient pathway toward accelerating the design and optimization of clean hydrogen energy systems.
Original/Review Paper
H.3.2.2. Computer vision
Mohammad Hossein Khosravi
Abstract
Document Image Quality Assessment (DIQA) is critical for ensuring the reliability of downstream applications such as Optical Character Recognition (OCR), digital archiving, and automated document workflows. In this paper, we propose a deep learning-based DIQA framework using a Siamese neural network ...
Read More
Document Image Quality Assessment (DIQA) is critical for ensuring the reliability of downstream applications such as Optical Character Recognition (OCR), digital archiving, and automated document workflows. In this paper, we propose a deep learning-based DIQA framework using a Siamese neural network architecture with an InceptionV3 backbone. Our model leverages a composite loss function that combines linear regression loss with a monotonic ranking constraint to jointly optimize for score-level accuracy and perceptual consistency. Unlike prior works that rely on handcrafted features or narrow degradation types, our approach generalizes across diverse distortions commonly observed in scanned and photographed documents. Experimental results on the SOC and SmartDoc-QA datasets demonstrate that the proposed model exhibits a strong correlation with OCR accuracy, achieving SROCC values of 0.952 and 0.873, respectively, and outperforming several state-of-the-art DIQA methods.
Methodologies
A.5. I/O and Data Communications
mansoor zeinali; Mohammad Ahmad Hadi
Abstract
This paper proposes an automatic modulation classification (AMC) framework that combines STFT spectrograms, a custom four-block ResNet, and Particle Swarm Optimization (PSO) for hyperparameter tuning. Its main contribution is the end-to-end integration of meta-heuristic optimization, systematic ablation ...
Read More
This paper proposes an automatic modulation classification (AMC) framework that combines STFT spectrograms, a custom four-block ResNet, and Particle Swarm Optimization (PSO) for hyperparameter tuning. Its main contribution is the end-to-end integration of meta-heuristic optimization, systematic ablation analysis, and explicit evaluation under both AWGN and Rayleigh fading channels, which has not been jointly addressed in earlier STFT-ResNet studies.The framework classifies six digital modulation schemes: ASK, PSK, FSK, QASK, QPSK, and QFSK. PSO was chosen instead of grid search and Bayesian optimization because it can efficiently handle the mixed discrete-continuous search space, including learning rate and batch size, without requiring gradient information. It also achieved convergence within only ten iterations.A balanced synthetic dataset of 6,000 samples was generated, with 1,000 samples per modulation class and 1,024 time-domain points per sample at a sampling frequency of 100 kHz. Although the dataset is synthetic, the authors acknowledge that validating the approach on over-the-air datasets such as RadioML 2016/2018 remains important future work.Five-fold cross-validation and ablation experiments show that the STFT representation improves accuracy by 9.6 percentage points, while PSO provides an additional statistically significant gain of 3 percentage points over a spectrogram-only baseline. The proposed model achieves 98.3% accuracy at 0 dB SNR in an AWGN channel across six tested SNR levels. Under Rayleigh fading, it reaches an unweighted average accuracy of 91.5%, improving from 81.2% at 0 dB to 97.5% at 25 dB, demonstrating strong robustness in noisy and fading conditions.
Technical Paper
H.3.2.6. Games and infotainment
Mohammadreza Mohammadnejad; Morteza Dorrigiv; Farzin Yaghmaee
Abstract
Research in recommender systems has largely relied on standardized datasets such as MovieLens, Amazon Reviews, and Last.fm. However, these datasets are unsuitable for in-game recommendations, particularly in Multiplayer Online Battle Arenas (MOBAs), due to the sequential, team-based, and adversarial ...
Read More
Research in recommender systems has largely relied on standardized datasets such as MovieLens, Amazon Reviews, and Last.fm. However, these datasets are unsuitable for in-game recommendations, particularly in Multiplayer Online Battle Arenas (MOBAs), due to the sequential, team-based, and adversarial nature of gameplay. To identify essential characteristics for in-game recommendation datasets, we perform a cross-domain analysis of widely used recommendation datasets, evaluating their structural and distributional properties, including interaction space, matrix shape, sparsity, and Gini-based feature–shape diversity. Building on these insights, we curate DOTA-Draft, a research-ready dataset from raw professional Dota 2 matches, encoding sequential pick/ban states, patch versions, and match outcomes. Using this dataset, we conduct top-k drafting recommendation tasks and provide baseline results with Bayesian Personalized Ranking (BPR) and GRU4Rec. To facilitate adoption, DOTA-Draft is packaged in a RecBole-compatible format. This work establishes principled benchmarks for in-game recommendation, demonstrates the inadequacy of traditional user–item paradigms in dynamic, adversarial environments, and provides a foundation for developing models that account for sequential, multi-agent decision-making.
Original/Review Paper
H.5. Image Processing and Computer Vision
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
F.2.7. Optimization
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.
Original/Review Paper
H.3.12. Distributed Artificial Intelligence
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.
Applied Article
H.3.8. Natural Language Processing
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
H.5. Image Processing and Computer Vision
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.
Original/Review Paper
H.6.5.10. Remote sensing
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
H.6.5.2. Computer vision
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.
Original/Review Paper
H.3. Artificial Intelligence
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 %.
Original/Review Paper
H.5.7. Segmentation
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
H.6.5.2. Computer vision
Mahdi Davari; Razieh Rastgoo
Abstract
Detecting driver distraction is critically important, as it remains a major contributor to road accidents and traffic-related injuries worldwide. This study introduces a novel hybrid deep learning model that integrates Spatio-Temporal Graph Convolutional Networks (ST-GCN) with a Transformer Encoder and ...
Read More
Detecting driver distraction is critically important, as it remains a major contributor to road accidents and traffic-related injuries worldwide. This study introduces a novel hybrid deep learning model that integrates Spatio-Temporal Graph Convolutional Networks (ST-GCN) with a Transformer Encoder and Attention mechanisms to effectively detect distracted driving behaviors. The ST-GCN component captures spatial and temporal dependencies in 3D skeletal motion data, modeling the dynamic body movements of the driver. Following this, a Transformer Encoder is employed to further refine temporal representations by leveraging global attention, allowing the model to understand long-range dependencies and subtle behavioral patterns over time. In addition, an Attention mechanism is applied to emphasize the most informative joints and time frames. To address class imbalance in the dataset, the model uses a focal loss function, which helps focus training on more difficult-to-classify examples. The proposed approach is validated on the 3D skeletal Drive&Act dataset, where it achieves a high accuracy of 97.47%, outperforming existing models, particularly under challenging conditions such as poor lighting and complex driving environments. The system demonstrates strong potential for real-time driver monitoring, offering an intelligent solution to enhance road safety and reduce accident risks through early detection of driver distraction.
)
