Optical character recognition is a rapidly evolving field within pattern recognition, enabling the automatic conversion of printed or handwritten text images into machine-readable formats. This technology plays a critical role across various sectors, including banking, healthcare, government, and education. While Optical character recognition systems encompass multiple stages such as text detection, segmentation, and post-processing, this article focuses on text recognition as a core and technically challenging component. In particular, we provide an in-depth review of recent advances driven by Transformer-based models, which have significantly pushed the state-of-the-art. To contextualize these advancements, a detailed comparative analysis of Transformer-based techniques is presented against earlier deep learning approaches, highlighting their respective limitations and the improvements introduced by Transformers, including parallel sequence processing, global context modeling, better handling of long-range dependencies, and enhanced robustness to irregular or noisy text layouts. We also examine widely used benchmark datasets in the literature and provide a detailed discussion of the performance achieved by recent state-of-the-art methods. Finally, this survey outlines open research challenges and potential future directions. It aims to serve as a comprehensive reference for both novice and experienced researchers by summarizing the latest developments in text recognition, including architectures, datasets, evaluation metrics, and practical considerations in model performance trade-offs and deployment.

Reconfiguring power distribution networks under harmonic load conditions presents a complex many-objective optimization challenge, where most objectives conflict. In high-dimensional objective spaces, population-based algorithms relying on Pareto dominance experience reduced selection pressure, limiting their reconfiguration effectiveness. To address this limitation, this paper introduces a fuzzy extension of Pareto dominance that enables more flexible comparison among solutions. By incorporating a tunable superiority threshold and nested dominance conditions between two individuals, the proposed method enhances selection pressure while maintaining compatibility with Pareto dominance. This adaptive approach allows for finer discrimination among mutually non-dominant solutions throughout the optimization process. The fuzzy dominance mechanism was embedded into two optimization algorithms-MOEAIGDNS and ToP-and evaluated using standard DTLZ and MaF benchmark problems and two customized distribution network scenarios with harmonic loads. Performance was measured using Hypervolume and Spacing indicators and statistically validated with the Mann-Whitney U test. Experimental results demonstrate that the proposed fuzzy dominance approach significantly improves optimization performance. Notably, ToP showed improvements in 84% of the cases based on Spacing, while MOEAIGDNS achieved superior Hypervolume results in 100% of the cases. Moreover, statistical analysis revealed that MOEAIGDNS with fuzzy dominance failed in only 2 % of all tests, underscoring the method's robustness. These results highlight the potential of fuzzy dominance to enhance many-objective optimization in complex, real-world applications.

Classification is one of the most important tasks in machine learning and data mining. Most of the classifiers are designed for data sets with equally distributed samples among the classes. Therefore, they encounter a problem with classifying imbalanced data in which one or more classes have much fewer samples than the others. Imbalanced data sets are prevalent in the real-world, so addressing this issue is of utmost importance. There have been many methods suggested to solve this problem showing promising results, a category of which is data-level methods being popular for their flexibility. In this paper, our goal is to review data-level methods comprehensively and categorize them from different perspectives. Also, to simplify doing future research in this field, most of the available benchmark imbalanced data sets, software, and toolboxes are introduced. Finally, existing challenges and future works are elaborated.

Farsi optical character recognition remains challenging due to the script’s cursive structure, positional glyph variations, and frequent diacritics. This study conducts a comparative evaluation of five foundational deep learning architectures widely used in OCR—two lightweight CRNN based models aimed at efficient deployment and three Transformer based models designed for advanced contextual modeling—to examine their suitability for the distinct characteristics of Farsi script. Performance was benchmarked on four publicly available datasets: Shotor and IDPL PFOD2 for printed text, and Iranshahr and Sadri for handwritten text, using word level accuracy, parameter count, and computational cost as evaluation criteria. CRNN based models achieved high accuracy on word level datasets—99.42% (Shotor), 97.08% (Iranshahr), 98.86% (Sadri)—while maintaining smaller model sizes and lower computational demands. However, their accuracy dropped to 78.49% on the larger and more diverse line level IDPL PFOD2 dataset. Transformer based models substantially narrowed this performance gap, exhibiting greater robustness to variations in font, style, and layout, with the best model reaching 92.81% on IDPL PFOD2. To the best of our knowledge, this work is among the first comprehensive comparative studies of lightweight CRNN and Transformer based architectures for Farsi OCR, encompassing both printed and handwritten scripts, and establishes a solid performance baseline for future research and deployment strategies.

Reinforcement learning, as a learning paradigm based on the interaction between an agent and its environment, facilitates the learning process and enhances the agent’s behavior through feedback derived from these interactions. However, these interactions pose challenges such as time costs, risks associated with operating in real-world environments, and the problem of missing data. In this research, to reduce agent-environment interactions, a method based on fuzzy reward granulation is proposed. This method forms fuzzy granules by evaluating the similarity among the members of the replay buffer. Each new interaction is either similar to one of these granules or is defined as the center of a new fuzzy granule. Fine-tuning the degree of fuzzy similarity strikes a balance between reducing interactions and exploring the environment. The evaluation of the proposed method is conducted by implementing it on various deep reinforcement learning algorithms, including DQN, DDQN, DuelingDQN, PPO, and SAC. These algorithms are employed to address 11 benchmark environments, covering both discrete and continuous action spaces, to comprehensively assess the method’s performance and adaptability. Subsequently, the Mann-Whitney U test is employed to assess the statistically significant differences in performance between agents trained using the algorithm enhanced with the proposed method and those trained with the baseline algorithm. The evaluation results indicate that the proposed method not only reduces the agent’s environmental interactions but also improves learning efficiency and effectiveness. This enhancement requires careful and precise calibration of the parameters introduced in the proposed method.

This work presents a pattern-aware framework for automatic license plate recognition (ALPR), designed to operate reliably across diverse plate layouts and challenging real-world conditions. The proposed system consists of a modern, high-precision detection network followed by a recognition stage that integrates a transformer-based vision model with an iterative language modelling mechanism. This unified recognition stage performs character identification and post-OCR refinement in a seamless process, learning the structural patterns and formatting rules specific to license plates without relying on explicit heuristic corrections or manual layout classification. Through this design, the system jointly optimizes visual and linguistic cues, enables iterative refinement to improve OCR accuracy under noise, distortion, and unconventional fonts, and achieves layout-independent recognition across multiple international datasets (IR-LPR, UFPR-ALPR, AOLP). Experimental results demonstrate superior accuracy and robustness compared to recent segmentation-free approaches, highlighting how embedding pattern analysis within the recognition stage bridges computer vision and language modelling for enhanced adaptability in intelligent transportation and surveillance applications.

Detecting small defects in power transmission lines is crucial for ensuring the safety and reliability of power grids. Unmanned Aerial Vehicle (UAV) imagery can be utilized for this purpose; however, the small size of defects and the low resolution of images make this task challenging. In this paper, we introduce a novel model that combines super-resolution and object detection techniques to address this issue. Our approach employs a Power Transmission Lines Single Image Super-Resolution Generative Adversarial Network (PTSRGAN) for super-resolution, and a new architecture for object detection that includes the HorNet backbone, a Super-Resolution based Path Aggregation Feature Pyramid Network (SR-PAFPN) neck, and a You Only Look Once-X (YOLOX) decoupled head. The SR-PAFPN neck enhances feature quality and diversity, particularly for small defects, by integrating feature super-resolution during training. To further improve the accuracy of small defect detection, our model is trained end-to-end, allowing the super-resolution model to receive feedback from the object detection model and adapt accordingly. Extensive experiments demonstrate the effectiveness and efficiency of our Power Transmission Lines Super Resolution Defect Detection (PTSRDet) method. Our model achieves a precision of 92.87% and a recall of 96.32%, processing each image in just 0.34 s. These results highlight the model’s capability to accurately detect small defects in power transmission lines, making it a valuable contribution to the field.

Outlier detection is an important issue in data mining, which has a wide range of applications in medicine, economics, video search, and credit card fraud detection. Many outlier detection methods have recently been developed. Most of the existing methods act based on the distance or density. Since each of these methods has its inherent disadvantage, we proposed a method which has the advantages of both distance-based and density-based methods. The proposed method is inspired by the basic idea that outliers are usually more distant neighbors to their nearest neighbors. The proposed method consists of three different parts. Each of these parts considers the distance, density, or location of objects, and finally we reach an optimal and efficient algorithm by combining these parts. Our algorithm is based on k nearest neighbor; in addition, we also use another kind of adaptive and extended neighborhood in order to provide more accurate results. Furthermore, the proposed method is robust and has little sensitivity to changes in parameter k. Numerical experiments and comparing with well-known algorithms are performed on both synthetic and real datasets in order to prove the efficiency and robustness of the proposed method.

Schizophrenia is a complex psychiatric disorder that disrupts cognition, emotions, and social behavior. Timely and accurate diagnosis is essential for effective treatment. Traditional diagnostic methods relying on clinical assessments have limitations, including subjectivity and time consumption. To address these challenges, there is increasing interest in utilizing neuroimaging techniques like electroencephalography (EEG) for schizophrenia diagnosis. EEG provides direct measures of brain activity and can reveal unique patterns associated with the disorder. This study proposes a novel approach that utilizes EEG signals to accurately diagnose schizophrenia, aiming to overcome the limitations of traditional methods. EEG data was collected from two groups: individuals diagnosed with schizophrenia and a healthy control group. During a visual task, EEG signals were recorded and underwent preprocessing to remove artifacts and noise. The data was segmented into non-overlapping time windows, and functional and time-domain features were extracted. These features were then used as inputs to a hybrid deep neural network specifically designed for EEG data. The primary objective of the network was to distinguish between healthy individuals and those with schizophrenia. The proposed approach was benchmarked against established methods such as support vector machines and k-nearest neighbors, and it demonstrated superior performance. The evaluation was carried out using a robust k-fold cross-validation approach. Various performance metrics, including accuracy, sensitivity, and characteristic criterion, were used to assess the diagnostic accuracy and discriminative power of the network. The results showed that the hybrid deep neural network effectively identified individuals with schizophrenia. This study highlights the potential of EEG-based diagnostic approaches in accurately diagnosing schizophrenia and offers a promising avenue for reducing subjectivity and improving the efficiency of the diagnostic process. Future research should focus on expanding the dataset, investigating generalizability across different populations, and exploring potential clinical applications in real-world settings.

The rapid growth of video data has resulted in an increasing need for surveillance and violence detection systems. Although such events occur less frequently than normal activities, developing automated video surveillance systems for violence detection has become essential to minimize labor and time waste. Detecting violent activity in videos is a challenging task due to the variability and diversity of violent behavior, which can involve a wide range of actions, motions, and interactions between people and objects. Currently, researchers employ deep learning models to detect violent behaviors. In fact, a large number of deep learning approaches are based on extracting spatio-temporal information from a video by exploiting a 3D Convolutional Neural Network (CNN). Despite their success, these techniques require a lot more parameters than 2D CNNs and have high computational complexity. Therefore, we focus on exploiting a 2D CNN to encode spatio-temporal information. Actually, statistical features of the optical flow changes are used to give this ability to a 2D CNN. These features are designed to make attention to regions of a video clip with much more motion. Accordingly, the optical flow of an input video is calculated. To determine meaningful changes in the optical flow, the optical flow magnitude of a current frame is compared with its predecessor. After that, statistical features of these changes are extracted to summarize a video clip to a 2D template, which feeds a 2D CNN. Experimental results on four benchmark datasets observe that the suggested strategy outperforms baseline ones. In particular, we make a better estimation of the spatio-temporal features in a video by shortening a video clip into a 2D template.

Mineral exploration is highly dependent on an accurate lithological map of a study area, which provides comprehensive information on geologic features for exploration target zones. Nowadays, the processing of hyperspectral image data for lithological mapping and mineral exploration using machine learning (ML) algorithms has greatly developed. The recently launched Italian hyperspectral sensor 'PRecursore IperSpettrale della Missione Applicativa (PRISMA)' offers an excellent capability for mineral detection and object classification with superior accuracy and efficiency for lithological mapping and mineral exploration. In this study, the performance of the support vector machine (SVM) algorithm was evaluated for processing PRISMA datasets to generate lithological maps of the Sar Cheshmeh porphyritic copper ore deposit in the Sahand-Bazman magmatic arc in central Iran. Three different SVM kernels, namely linear (LSVM), quadratic (QSVM) and cubic (CSVM), were comparatively evaluated for data classification in lithological mapping. The SVM classifiers were trained on the basis of prior knowledge from previous studies and field surveys. Approximately 5000 pixels from 14 different classes were used for training. There was a large misclassification between granodiorites and altered granodiorites in the LSVM result (78.3% accuracy for altered granodiorites), but this was greatly reduced in the QSVM and CSVM methods (with 96.1% and 99.1% accuracy, respectively). A significant improvement in classification was also seen for the vegetation, mine pits and Razak volcanism classes (with varying accuracy values). It is noteworthy that nine of the 14 classes had less than 400 training pixels and only one class had more than 1000 pixels used for training, indicating the power of ML for such studies. LSVM was the best method for mapping dacites with maximum accuracy (100%), but this accuracy was slightly lower for QSVM and CSVM (both had 97.9% accuracy). The results show that the LSVM, QSVM and CSVM methods achieved an accuracy of 80.22%, 85.81% and 86.05%, respectively, in the final classification. This study advocates the optimal SVM classifier (CSVM classifier) using PRISMA hyperspectral images for accurate lithological mapping for mineral exploration in metallogenic provinces.

The pelletizing process is a key stage in the steel production chain, during which iron ore concentrate is transformed into pellets suitable for use in blast furnaces. The induration (firing) stage within the furnace plays a critical role in determining product quality, as the distribution of temperature and pressure directly affects the physical and mechanical properties of the pellets. The objective of this study is to develop a data-driven model to predict the behavior of the pelletizing furnace at Golgohar Mining and Industrial Company based on real operational data. To this end, process input and output variables were collected and preprocessed, and feature selection was performed using three methods: F-score, mutual information, and Pearson correlation coefficient. Subsequently, Multiple Linear Regression, random forest, k-nearest neighbors, and multilayer perceptron neural network models were trained within a multiple-input–multiple-output (MIMO) framework to simultaneously predict 36 temperature variables and 31 pressure variables. The results indicated that the mutual information-based feature selection method improved the performance of nonlinear models, while the multilayer perceptron achieved the highest accuracy, with an average coefficient of determination of 91.50% for temperature and 89.62% for pressure. These findings demonstrate that data-driven learning approaches can accurately model the complex behavior of the furnace and provide a suitable foundation for designing intelligent control systems and optimizing firing conditions

Portfolio management involves choosing and actively overseeing various investment assets to meet an investor’s long-term financial goals, considering their risk tolerance and desired return potential. Traditional methods, like mean–variance analysis, often lack the flexibility needed to navigate the complexities of today’s financial markets. Recently, Deep Reinforcement Learning (DRL) has emerged as a promising approach, enabling continuous adjustments to investment strategies based on market feedback without explicit price predictions. This paper presents a comprehensive literature review of DRL applications in portfolio management, aimed at finance researchers, data scientists, AI experts, FinTech engineers, and students seeking advanced portfolio optimization methodologies. We also conducted an experimental study to evaluate five DRL algorithms—Advantage Actor-Critic (A2C), Deep Deterministic Policy Gradient (DDPG), Proximal Policy Optimization (PPO), Soft Actor-Critic (SAC), and Twin Delayed DDPG (TD3)—in managing a portfolio of 30 Dow Jones Industrial Average (DJIA) stocks. Their performance is compared with the DJIA index and traditional strategies, demonstrating DRL’s potential to improve portfolio outcomes while effectively managing risk.

Colloids in saturated porous media, such as soil and aquifers, play a critical role in the transport of nutrients, pollutants, and microorganisms. Their movement can influence the quality of groundwater and the effectiveness of filtration systems. Detecting colloids in these environments is essential for understanding contaminant spread, predicting soil and groundwater behavior, and managing water resources. Accurate detection helps in designing remediation strategies and ensures the safe use of natural resources, particularly in environmental engineering and hydrogeology. In this paper, we apply an artificial intelligence approach with the help of deep learning to detect colloids, which is a prerequisite for subsequent steps in porous media research. Since colloids are tiny particles and do not have enough information to identify, firstly we use an image processing technique called the dilation operation to improve distinguishing features of colloids for the detection process. This operation leads to achieving more accurate results for the detection of tiny colloids. Then, we propose a lightweight deep convolutional neural network to detect colloids automatically without the requirement for manual analysis. In our experiments, Precision, Recall, F-measure, and TCR metrics are employed for assessment. The experimental results show the efficiency and effectiveness of the proposed approach compared to six image processing methods in the detection process of colloids.

Over recent decades, natural and artificial colloids, as well as nanoparticles, have been increasingly used in various applications. Consequently, with this rising consumption, surface and subsurface environments are more exposed to these particles. The presence of these particles and the colloid-facilitated transport of microorganisms, the interactions between dissolved contaminants and mobile colloids in porous media, and the fate and transport of colloids through groundwater—one of the primary sources of water supply for human societies—have attracted extensive research. This study investigates the performance of several image processing methods in the field of colloid detection, which is a prerequisite for the subsequent steps in porous media research. We employed four different categories of image processing approaches on microscopy images—segmentation-based methods, background-detection-based methods, filter-based methods, and morphology-based methods—to conduct the detection process of colloids. Eight methods were applied and subsequently analyzed in terms of their drawbacks and advantages to determine the best ones in this domain. Finally, we proposed an ensemble approach that leverages the strengths of the three best methods using a majority vote to detect colloids more accurately. In experiments, Precision, Recall, F-measure, and TCR criteria were considered as evaluation tools. Experimental results demonstrate the high accuracy of image processing methods in recognizing colloids. Among all these methods, morphology-based methods were the most successful, achieving the best detection performance and improving the limited distinguishing features of small colloids. Moreover, our ensemble approach, achieving perfect scores across all evaluation criteria, highlights its superiority compared with other detection methods.

While crucial for document forensics and security, detecting Persian signatures in real-world scenarios poses a considerable challenge due to the distinctive features of Persian signatures—characterized by complex combinations of shapes and letters instead of conventional names—along with issues like noise, cluttered backgrounds, and more. Furthermore, the scarcity of annotated datasets complicates signature detection in natural settings. This paper addresses these challenges by introducing a new dataset for Persian documents, merging publicly available signatures with diverse backgrounds to create semi-realistic images reflecting real-world conditions. Additionally, the proposed method leverages the YOLOv5 architecture for signature detection. The experimental findings illustrate that, under various lighting conditions, background complexities, and signature distortions, this approach effectively and accurately detects Persian signatures.

Geological mapping of exposed geologic units of Earth surface is a common and important activity for geologists. This process is the first step of geological prospecting. Remote sensing can provide useful driven data for further studies and also it reduces the time and cost of this process. Sometimes it is possible that some lithologies have similar spectral responses while they have different surficial texture properties. Geological units of the Earth surface are more separable by including their textural properties along with their spectral behaviour in remote sensing so authors used spectral-Radar data integration with novel idea which is named Radar data resultant vector in this study. In this paper, two different neural network methods (Neural Pattern Recognition and Neural Net Fitting) were implemented in Matlab environment for lithological classification using two different input datasets, namely (1) only multispectral data and (2) integrated Radarmultispectral data. The reason was to evaluate the performance of Spectral-Radar fused data in lithological classification in comparison with the spectral data alone. The results show that integrated Radar-multispectral data results in better classification of lithological units due to the integration of surficial textural parameters and spectral responses of such surficial features. The results also showed that the Neural Pattern Recognition method (NPRTool) performed better than Neural Net Fitting (NFTool) method. The results further show that, among the three different algorithms of the Neural Net Fitting method (i.e., Levenberg-Marquardt, Bayesian Regularization, and Scaled Conjugate Gradient), the Levenberg-Marquardt performed best.

In today's digital age, there's a growing demand for a dependable system that can recognize handwritten text. This demand comes from various industries as well as individual users, highlighting the importance of a system that can quickly convert, store, and enhance the accessibility of handwritten documents. Encoder-decoder models have proven their effectiveness in various sequence learning tasks, such as machine translation, image captioning, action recognition, and optical character recognition. However, these models have encountered difficulties in achieving competitive results in the specific context of Farsi handwritten text recognition. This paper introduces an innovative lightweight attention-driven encoder-decoder model designed to address the challenges associated with Farsi handwritten text recognition. To evaluate the proposed model, two well-known Farsi handwritten text datasets, Sadri and Iranshahr, are utilized. Additionally, to potentially enhance recognition accuracy, two different convolutional backbones-ResNet34 and MobileNetV3-are explored. Through experimental analysis, our best-performing model surpasses existing CNN-based and CRNN-based methods, achieving recognition accuracies of 94.79% on the Iranshahr dataset and 98.99% on the Sadri dataset. Additionally, our model offers faster inference speeds and a lower computational burden compared to recent Transformer-based approaches. These results demonstrate the effectiveness of our attention-driven encoder-decoder model in addressing the challenges of lightweight Farsi handwritten text recognition.

Outlier detection has garnered considerable attention in recent years due to its wide-ranging applications across various research domains. This surge in interest has led to the development of numerous detection techniques, predominantly based on distance or density metrics. A notable limitation of these existing methods is their reliance on parameter adjustments, significantly affecting the outcome. Additionally, these methods exhibit intrinsic flaws: distance-based approaches struggle with clusters with varying local densities, while density-based methods fail to identify patterns within low-density areas. Moreover, most prior techniques are adept at identifying only one kind of outlier—local, global, or group of outliers. Addressing these challenges, we introduce the Adaptive Radius Density-Based Outlier Detection (ARDOD) method, which departs from the traditional parameter-dependent approach. ARDOD is a novel parameter-free algorithm that dynamically determines the necessary parameters based on the data distribution within the feature space. This innovative method demonstrates robust performance in detecting all three categories of outliers. The efficacy and superior performance of ARDOD are validated through an extensive experimental analysis involving various synthetic and real-world datasets. This analysis showcases ARDOD's advantages over seven established methods: Local Outlier Factor (LOF), Angle-Based Outlier Detection (ABOD), Robust Distance-Based Outlier Score (RDOS), Directed density ratio Changing Rate-based outlier detection (DCROD), Empirical-Cumulative-distribution-based Outlier Detection(ECOD), mean-shift outlier detector(MOD +),and Local–Global Outlier Detection (LGOD), underscoring its potential as a versatile tool in outlier detection research.

One of the most challenging problems in machine learning is the classification of imbalanced data. This problem arises when the samples of data are distributed unevenly among the classes, such that compared to one class (the majority or negative class), the other class (the minority or positive class) has far fewer samples. The classical classifiers are inappropriate to classify data sets of this nature. To address these classifiers’ shortcoming in class imbalance situations, we present a novel clustering-based over-sampling technique in this paper. First, the $k$ -means clustering algorithm is used to cluster the minority class samples. Then, sparse clusters including fewer samples are chosen. Finally, we use the nearest neighbor of each cluster center to create synthetic samples for the minority class. Also, to select clusters based on probabilities, we apply the roulette wheel selection operator during over-sampling. The C4.5 decision tree classifier is utilized in our experiments, and AUC and $F$ -measure metrics are considered to evaluate methods. According to the results, our method outperforms six other methods over fifteen imbalanced data sets.