Accurate diagnosis of neurological disorders is crucial for making effective and early clinical decisions and managing long-term disease outcomes. Recent advances in deep learning, particularly Vision Transformer (ViT) models, have demonstrated strong potential in modeling global contextual dependencies in neuroimaging data. Motivated by these developments, this systematic literature review provides a comprehensive analysis of recent advancements in ViT models for neurological disorders. The study presents a technical taxonomy of ViT models categorized into pure, hybrid, hierarchical, and lightweight architectures by employing an adapted PRISMA method. Among the selected studies, 54% are categorized as hybrid ViT, while 22%, 17%, and 7% are classified as hierarchical, pure, and lightweight ViTs, respectively. Additionally, the commonly used datasets and various data types for diagnosing neurological disorders are described. Neuroimaging emerged as the most common data type, employed in 78% of the studies, followed by signal (12%), behavioral (4%), clinical (4%), and gene data (2%). This study also analyses the preprocessing and postprocessing methods employed for neuroimaging data within the selected studies. Finally, the study discusses the identified challenges and open issues, suggesting future research directions concerning the implementation of diverse ViT architectures in the neurological domain.

Accurate river streamflow forecasting is vital for water resources management, flood control, hydrological and water sciences. Robust prediction models are required due to nonlinearity and complexity of river streamflow process. To address this, different deep learning (DL) methods, including Transformer-based architectures, have been recently developed. This study evaluates the performance of six diverse variants of DL techniques in multi-step ahead forecasting of the daily river streamflows at two hydrometric river stations (10AA001 and 10ED002) located on the Liard River, Canada. The developed models included five baseline models: Frequency Enhanced Decomposed Transformer (FEDformer), Informer, Long Short-Term Memory (LSTM), Transformer, and Inverted Transformer (iTransformer). In addition, a hybrid scheme named FEDformer-LSTM was proposed by integrating the FEDformer and LSTM. The obtained results demonstrated that the proposed FEDformer-LSTM hybrid model not only outperformed the baseline FEDformer but also showed the best performance among all the models developed, confirming its reliable capability for multi-step ahead forecasting of the daily river streamflows at the studied stations. An attempt was finally made to improve the models’ performances through a preprocessing method, namely Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN). The findings revealed that the hybrid CEEMDAN-DL frameworks generally achieved better outcomes compared to the relevant baseline DL models. The best river streamflow forecasts were achieved through CEEMDAN-LSTM at both the stations at t + 1 lead time (10AA001 Station: MAE = 6.4435 m3/s, MSE = 251.5336 (m3/s)2, RMSE = 15.8598 m3/s, R2 = 0.9989, NSE = 0.9988, KGE = 0.9911, PBIAS = −0.0637%; 10ED002 Station: MAE = 55.7703 m3/s, MSE = 9249.1674 (m3/s)2, RMSE = 96.1726 m3/s, R2 = 0.9994, NSE = 0.999, KGE = 0.978, PBIAS = 1.7122%). Therefore, the models developed in this study can be proposed as efficient intelligent tools for accurate river streamflow forecasting for optimal water resources management and allocation.

Accurate and efficient surface defect segmentation is essential in assuring the quality and efficiency of products in the industrial sector. While current Convolutional Neural Networks (CNNs) based on encoder-decoder architecture have achieved promising results for defect detection, they often struggle to capture long-range dependencies. In contrast, Transformer-based models recently demonstrated strong capabilities in global feature extraction but are limited to local features for the refined defect location. To address these limitations, this paper proposes a Label-Aware CNN-Transformer Network (LACTNet), which integrates explicit label-aware supervision and adaptive feature fusion for surface defect segmentation. The proposed model utilizes a dual-encoder architecture, employing convolutional layers to extract local features and the Swin Transformer to extract global features for surface defect segmentation.An adaptive feature fusion module (AFFM) is introduced to effectively aggregate the extracted features at each step of the encoder layers. Moreover, a multi-stage Label-Aware Module (LAM) is proposed to enhance the learning process by adding supervision in the decoder layers. We analyzed the performance of the proposed model using three public benchmark datasets. Based on our results, Dice scores of the LACTNet are as follows: 0.9328, 0.8339, 0.9757 for SD-saliency-900, NRSD-MN, and DAGM2007, respectively.

Microbial biomass carbon (MBC) is a key indicator of soil quality, and comprehensive knowledge of MBC is essential for sustainable environmental management and effective monitoring of soil quality changes. However, directly measuring MBC is time-consuming and costly. To overcome these limitations, artificial intelligence (AI) models offer a promising alternative by analyzing easily measurable soil properties. In this study, soil parameters such as organic carbon (OC), clay content, bicarbonate anion (HCO3−), and total nitrogen (TN) concentrations were used as input features for AI models to predict MBC. A novel hybrid model combining a multilayer perceptron artificial neural network (MLP) with the Hunger Games Search (HGS) optimization algorithm, referred to as the Hybrid HGS–MLP, was developed. The predictive performance of this model was evaluated against several conventional models, including standalone MLP, Gene Expression Programming (GEP), Adaptive Neuro-Fuzzy Inference System (ANFIS), Multivariate Adaptive Regression Splines (MARS), and Multiple Linear Regression (MLR). Model accuracy and reliability were assessed using the coefficient of determination (R2), root mean square error (RMSE), and mean absolute error (MAE). Results showed that all AI-based models (Hybrid HGS–MLP, standalone MLP, and GEP) achieved an average relative error of less than 1

The conventional wind forecasting methods often struggle to handle the non-stationary and inconsistent wind patterns. This paper presents a hybrid method of Empirical Wavelet Transform (EWT) and Deep Reinforcement Learning (DRL) for wind speed modeling to overcome the forecasting challenges. The EWT method transforms the original wind speed series into several independent modes and a residual series. In addition, the DRL method is utilised to optimise the weights associated with three distinct supervised deep learning models, i.e., Long Short-Term Memory (LSTM), Convolutional Neural Networks with LSTM (CNN-LSTM), and CNN with Gated Recurrent Units (CNN-GRU). The performance of the proposed EWT-DRL is evaluated against deep learning models, including LSTM, CNN-LSTM, CNN-GRU, and their coupling with EWT. The combination of EWT and the DRL (EWT-DRL) method achieves a Mean Absolute Error (MAE) of 0.151, a Mean Squared Error (MSE) of 0.060, a Root Mean Squared Error (RMSE) of 0.192, and a correlation coefficient (R) of 0.9913. These results indicate the effectiveness of EWT-DRL in improving accuracy for wind speed modeling.

Solar energy is one of the renewable and clean energy sources. Accurate solar radiation (SR) estimates are therefore needed in solar energy applications. Firstly, two deep learning models, including gated recurrent unit (GRU) and long short-term memory (LSTM), were developed in this study. Next, a data pre-processing technique named multivariate variational mode decomposition (MVMD) was used to construct the MVMD-GRU and MVMD-LSTM hybrid models. To better test the performance of proposed simple and hybrid models, four stations located in the Illinois State of the USA (i.e., Dixon Springs, Fairfield, Rend Lake, and Carbondale) were considered as the study sites. Whole the simple and hybrid models were established under two different strategies, i.e., local and external. In the local strategy, SR of each location was estimated using the minimum and maximum air temperatures from the same station. While, minimum and maximum air temperatures as well as SR data from the nearby station were utilized in external strategy to estimate SR time series of any target site. Root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) metrics were used when evaluating the models performances. The overall results revealed that the proposed MVMD-GRU and MVMD-LSTM hybrid models illustrated better SR estimates compared to the simple GRU and LSTM in both the local and external strategies. The values of error metrics obtained for the superior hybrid models (i.e., MVMD-LSTM) during the testing period were as: RMSE = 2.532 MJ/m2.day, MAE = 1.921 MJ/m2.day, R2 = 0.916 at Dixon Springs; RMSE = 2.476 MJ/m2.day, MAE = 1.878 MJ/m2.day, R2 = 0.921 at Fairfield; RMSE = 2.359 MJ/m2.day, MAE = 1.780 MJ/m2.day, R2 = 0.924 at Rend Lake; RMSE = 2.576 MJ/m2.day, MAE = 1.941 MJ/m2.day, R2 = 0.914 at Carbondale. Therefore, the coupled models proposed in this study can be possibly recommended as suitable alternatives to the simple deep learning models with a reliable precision in estimating SR time series.

Most energy exchanges take place through the building skin. The skin characteristics play a decisive role in the extent of these exchanges, but they are somewhat more varied in the double skin façade (DSF). Among these characteristics, cavity segmentation has a noticeable effect on the implementation of the DSF in different directions during the hot and cold seasons. The aim of this study was to investigate the role of DSF segmentation in energy consumption and natural ventilation of high-rise buildings in hot and dry climates. This study used DesignBuilder software to study sixty-four segmentation component scenarios in an eight-story residential building in Isfahan, Iran. Moreover, a proposed hybrid model utilizing the hybridization of the Hunger Game Search and Gradient Boosting (HGS-GB) algorithm was employed to estimate energy consumption in various scenarios involving lighting, heating, cooling, and total scenarios. The available outcomes revealed that the HGS-GB model had a better performance in comparison with other individual models, such as Gradient Boosting (GB), Random Forest (RF), and K-nearest neighbors (KNN). The R2 values for lighting, heating, cooling, and total energy estimation were 0.9993, 0.9958, 0.9991, and 0.9922, respectively. The findings of this study suggest the significance of DSF segmentation in energy consumption and natural ventilation in high-rise buildings in hot and dry climates.

Monitoring the quality of river water is of fundamental importance and needs to be taken into consideration when it comes to the research into the hydrological field. In this context, the concentration of the dissolved oxygen (DO) is one of the most significant indicators of the quality of river water. The current study aimed to estimate the minimum, maximum, and mean DO concentrations (DO min, DO max, DO mean) at a gauging station located on Tualatin River, United States. To that end, four machine learning models, such as support vector regression (SVR), multi-layer perceptron (MLP), random forest (RF), and gradient boosting (GB) were established. Root mean square error (RMSE), mean absolute error (MAE), coefficient of correlation (R), and Nash-Sutcliffe efficiency (NSE) metrics were employed to better assess the accuracies of these models. The modeling results demonstrated that the SVR and MLP surpassed the RF and GB models. Despite this, the SVR was concluded to be the best-performing method when used to estimate the DO min, DO max, and DO mean. The best error statistics in the testing phase were related to the SVR model with full (four) inputs to estimate DO mean concentration (RMSE = 0.663 mg/l, MAE = 0.508 mg/l, R = 0.945, NSE = 0.875). Finally, the explainability of the superior models (i.e. SVR models) was conducted using SHapley Additive exPlanations (SHAP) for the first time to estimate DO concentration. In fact, evaluating the explainability of machine learning models can provide useful information about the impact of each of the input estimators used in the procedure of models development. It was concluded that the specific conductance (SC) and followed by water temperature (WT) could provide the most contributions for estimating the DO min, DO max, and DO mean concentrations.

Background Computerized posturography obtained in standing conditions has been applied to classify fall risk for older adults or disease groups. Combining machine learning (ML) approaches is superior to traditional regression analysis for its ability to handle complex data regarding its characteristics of being high-dimensional, non-linear, and highly correlated. The study goal was to use ML algorithms to classify fall risks in community-dwelling older adults with the aid of an explainable artificial intelligence (XAI) approach to increase interpretability. Methods A total of 215 participants were included for analysis. The input information included personal metrics and posturographic parameters obtained from a tracker-based posturography of four standing postures. Two classification criteria were used: with a previous history of falls and the timed-up-and-go (TUG) test. We used three meta-heuristic methods for feature selection to handle the large numbers of parameters and improve efficacy, and the SHapley Additive exPlanations (SHAP) method was used to display the weights of the selected features on the model. Results The results showed that posturographic parameters could classify the participants with TUG scores higher or lower than 10 s but were less effective in classifying fall risk according to previous fall history. Feature selections improved the accuracy with the TUG as the classification label, and the Slime Mould Algorithm had the best performance (accuracy: 0.72 to 0.77, area under the curve: 0.80 to 0.90). In contrast, feature selection did not improve the model performance significantly with the previous fall history as a classification label. The SHAP values also helped to display the importance of different features in the model. Conclusion Posturographic parameters in standing can be used to classify fall risks with high accuracy based on the TUG scores in community-dwelling older adults. Using feature selection improves the model’s performance. The results highlight the potential utility of ML algorithms and XAI to provide guidance for developing more robust and accurate fall classification models. Trial registration Not applicable

Global adoption of wind energy continues to increase, while improving the efficiency of turbine settings requires reliable wind speed (WS) models. The latest models rely on artificial intelligence (AI) optimizations which constructs tests on a range of novel hybrid models to examine the reliability. Gradient Boosting (GB), Random Forest (RF), and Long Short-Term Memory (LSTM) are used in new combinations for data pre-processing. A Time Varying Filter-based Empirical Mode Decomposition (TVFEMD) model is coupled with the GB and LSTM standalone models, to create TVFEMD-GB and TVFEMD-LSTM hybrids, which are run in competition with each other. Eventually, a preferred hybrid form is established, simultaneous hybridization of TVFEMD with GB and LSTM. This study is the first to hybridize these fundamental systems, and create a TVFEMD-GB-LSTM model that can forecast WS. This study finds that the novel hybrid models exhibit superior performance to standalone GB and LSTM models, opening the pathway to alternative WS prediction techniques.

Detecting surface defects plays a crucial role in ensuring the quality, functionality, and security of the production process. Traditional image processing techniques and machine learning models rely on manual analysis and feature extraction for specific vision inspection tasks. Deep learning approaches, which can automatically extract features from images, have demonstrated outstanding performance in computer vision tasks, including detecting surface defects. Motivated by this consideration, a Systematic Literature Review (SLR) method is employed for the comprehensive analysis of studies published between 2020 and 2023 in the field of deep learning-based surface defect detection applications in industrial products. The study provides a technical taxonomy for deep learning models according to the content of current studies through the SLR process, including Convolutional Neural Networks (CNN), encoder–decoder models, pyramid network models, Generative Adversarial Networks (GAN), attention-based models, and other models for surface defect detection. Then, the commonly used datasets for surface defect detection are discussed, and a comparative analysis of deep learning models’ performance is provided. Our comparative analysis reveals that pyramid network models and CNN models are the most frequently used deep learning models for surface defect detection. These models yield reasonable results in surface defect detection due to their exceptional feature extraction capabilities. Finally, some hints for addressing future research directions and identifying open issues in surface defect detection applications are presented.

Precise detection and localization of the Endotracheal tube (ETT) is essential for patients receiving chest radiographs. A robust deep learning model based on U-Net++ architecture is presented for accurate segmentation and localization of the ETT. Different types of loss functions related to distribution and region-based loss functions are evaluated in this paper. Then, various integrations of distribution and region-based loss functions (compound loss function) have been applied to obtain the best intersection over union (IOU) for ETT segmentation. The main purpose of the presented study is to maximize IOU for ETT segmentation, and also minimize the error range that needs to be considered during calculation of distance between the real and predicted ETT by obtaining the best integration of the distribution and region loss functions (compound loss function) for training the U-Net++ model. We analyzed the performance of our model using chest radiograph from the Dalin Tzu Chi Hospital in Taiwan. The results of applying the integration of distribution-based and region-based loss functions on the Dalin Tzu Chi Hospital dataset show enhanced segmentation performance compared to other single loss functions. Moreover, according to the obtained results, the combination of Matthews Correlation Coefficient (MCC) and Tversky loss functions, which is a hybrid loss function, has shown the best performance on ETT segmentation based on its ground truth with an IOU value of 0.8683.

This paper investigates the applications of explainable AI (XAI) in healthcare, which aims to provide transparency, fairness, accuracy, generality, and comprehensibility to the results obtained from AI and ML algorithms in decision-making systems. The black box nature of AI and ML systems has remained a challenge in healthcare, and interpretable AI and ML techniques can potentially address this issue. Here we critically review previous studies related to the interpretability of ML and AI methods in medical systems. Descriptions of various types of XAI methods such as layer-wise relevance propagation (LRP), Uniform Manifold Approximation and Projection (UMAP), Local Interpretable Model-agnostic Explanations (LIME), SHapley Additive exPlanations (SHAP), ANCHOR, contextual importance and utility (CIU), Training calibration-based explainers (TraCE), Gradient-weighted Class Activation Mapping (Grad-CAM), t-distributed Stochastic Neighbor Embedding (t-SNE), NeuroXAI, Explainable Cumulative Fuzzy Class Membership Criterion (X-CFCMC) along with the diseases which can be explained through these methods are provided throughout the paper. The paper also discusses how AI and ML technologies can transform healthcare services. The usability and reliability of the presented methods are summarized, including studies on the usability and reliability of XGBoost for mediastinal cysts and tumors, a 3D brain tumor segmentation network, and the TraCE method for medical image analysis. Overall, this paper aims to contribute to the growing field of XAI in healthcare and provide insights for researchers, practitioners, and decision-makers in the healthcare industry. Finally, we discuss the performance of XAI methods applied in medical health care systems. It is also needed to mention that a brief implemented method is provided in the methodology section.

River discharge represents critical hydrological data that can be used to monitor the hydrological status of a river basin. The objective of this study was to forecast the monthly river discharge time-series of two gauging hydrometric sites (USGS 06054500 and USGS 06090800) located on the Missouri River, USA. The forecast was performed using two machine learning models based on extreme gradient boosting (XGB) and K-nearest neighbors (KNN). XGB outperformed the KNN framework in forecasting the river flow. Subsequently, wavelet (W) analysis was incorporated to develop the hybrid W-XGB and W-KNN approaches. Finally, two novel hybrid models were established through the hybridization of XGB and the Lévy–Jaya optimization algorithm (LJA) and simultaneous integration of the wavelet analysis and LJA with the XGB, i.e., XGB-LJA and W-XGB-LJA, respectively. The performances of the models were evaluated using the root mean square error (RMSE), mean absolute error (MAE), mean bias error (MBE), determination coefficient (R ), and Nash–Sutcliffe efficiency (NSE). In the test phase, the best discharge forecasts at USGS 06054500 and USGS 06090800 were obtained using the hybrid WXGB2-LJA (RMSE = 41.303 m /s, MAE = 28.752 m /s, MBE = 3.377 m /s, R = 0.819, NSE = 0.800) and W-XGB4-LJA (RMSE = 39.310 m /s, MAE = 26.804 m /s, MBE = 1.489 m3/s, R = 0.897, NSE = 0.885), respectively.

With urbanization and increasing consumption, there is a growing need to prioritize sustainable development across various industries. Particularly, sustainable development is hindered by air pollution, which poses a threat to both living organisms and the environment. The emission of combustion gases containing particulate matter (PM 2.5) during human and social activities is a major cause of air pollution. To mitigate health risks, it is crucial to have accurate and reliable methods for forecasting PM 2.5 levels. In this study, we propose a novel approach that combines support vector machine (SVM) and long short-term memory (LSTM) with complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to forecast PM 2.5 concentrations. The methodology involves extracting Intrinsic mode function (IMF) components through CEEMDAN and subsequently applying different regression models (SVM and LSTM) to forecast each component. The Naive Evolution algorithm is employed to determine the optimal parameters for combining CEEMDAN, SVM, and LSTM. Daily PM 2.5 concentrations in Kaohsiung, Taiwan from 2019 to 2021 were collected to train models and evaluate their performance. The performance of the proposed model is evaluated using metrics such as mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and coefficient of determination (R2) for each district. Overall, our proposed model demonstrates superior performance in terms of MAE (1.858), MSE (7.2449), RMSE (2.6682), and (0.9169) values compared to other methods for 1-day ahead PM 2.5 forecasting. Furthermore, our proposed model also achieves the best performance in forecasting PM 2.5 for 3- and 7-day ahead predictions.

Dictionary learning was introduced for sparse image representation. Today, it is a cornerstone of image classification. We propose a novel dictionary learning method to recognise images of handwritten numbers. Our focus is to maximise the sparse-representation and discrimination power of the class-specific dictionaries. We, for the first time, adopt a new feature space, i.e., histogram of oriented gradients (HOG), to generate dictionary columns (atoms). The HOG features robustly describe fine details of hand-writings. We design an objective function followed by a minimisation technique to simultaneously incorporate these features. The proposed cost function benefits from a novel class-label penalty term constraining the associated minimisation approach to obtain class-specific dictionaries. The results of applying the proposed method on various handwritten image databases in three different languages show enhanced classification performance (∼98%) compared to other relevant methods. Moreover, we show that combination of HOG features with dictionary learning enhances the accuracy by 11% compared to when raw data are used. Finally, we demonstrate that our proposed approach achieves comparable results to that of existing deep learning models under the same experimental conditions but with a fraction of parameters.

River streamflow is an essential hydrological parameters for optimal water resource management. This study investigates models used to estimate monthly time-series river streamflow data at two hydrological stations in the USA (Heise and Irwin on Snake River, Idaho). Five diverse types of machine learning (ML) model were tested, support vector machine-radial basis function (SVM-RBF), SVM-Polynomial (SVM-Poly), decision tree (DT), gradient boosting (GB), random forest (RF), and long short-term memory (LSTM). These were trained and tested alongside a conventional multiple linear regression (MLR). To improve the estimation and model performance, hybrid models were designed by coupling the models with wavelet theory (W). The models performance was assessed using root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R-2), Nash-Sutcliffe efficiency (NSE), and Willmott's index (WI). A side-by-side performance assessment of the stand-alone and hybrid models revealed that the coupled models exhibit better estimates of monthly river streamflow relative to the stand-alone ones. The statistical parameter values for the best model (W-LSTM4) during the test phase was RMSE = 36.533 m(3)/s, MAE = 26.912 m(3)/s, R-2 = 0.947, NSE = 0.946, WI = 0.986 (Heise station), and RMSE = 33.378 m(3)/s, MAE = 24.562 m(3)/s, R-2 = 0.952, NSE = 0.951, WI = 0.987 (Irwin station).

Handwritten documents have been a valuable resource in human transactions for many years. Today, there is an immediate need for computer-based techniques to intelligently read and analyze such documents. Meanwhile, handwritten numerals are of particular importance due to their role in finance, business, post, etc. Although there exist many researches on English handwritten number recognition, the development of reliable recognition systems has been paid little attention for non-English scripts. In this chapter, an overview of the state-of-the-art on handwritten number recognition (with focus on non-English languages) is presented. Dictionary learning as a supervised learning technique, which has been recently shown great success in image classification problems, is introduced. We describe the ways one can design discriminative dictionaries for classification of handwritten numbers. The obtained dictionaries convey exclusive features of the associated numerals. In order to improve the classification performance of handwritten numbers using dictionary learning, two novel approaches are presented. First, an incoherence penalty is combined with the learning process to fine-tune the structure of the dictionaries learned for each class. Second, class label information is embedded into the learning process in order to produce class-specific weights which improve the discriminativity of the learned dictionaries. We further adopt a new feature space, that is, histogram of oriented gradients (HOG) to generate the dictionary atoms. HOG is a strong descriptor of most handwritten images especially those studied in this chapter. Four different handwritings, namely, Chinese, Persian, Arabic, as well as English are used to evaluate the performance of the proposed methods. We also present a convolutional neural network model to compare the performance of deep learning with that of dictionary learning for handwritten digits recognition. The obtained results and their comparisons with benchmark methods confirm the effectiveness and robustness of the proposed approaches for recognition of handwritten numbers.

Deep learning methods have become the key ingredient in the field of computer vision; in particular, convolutional neural networks (CNNs). Appropriating the network architecture and data pre-processing have significant impact on performance. This paper focuses on the classification of handwritten Chinese numbers. Firstly, we applied various methods of pre-processing to our collected image dataset. Secondly, we customised a CNN-based architecture with minimal number of layers and parameters specifically for the task. Experimental results showed that our proposed methods provides superior classification rate of 99.1%. Our results also show that the proposed method has competitive performance compared to smaller neural networks with fewer parameters, e.g. Squeezenet and deeper networks with a larger size and number of parameters, e.g., pre-trained GoogLeNet and MobileNetV2.

Dictionary learning is a cornerstone of image classification. We set out to address a longstanding challenge in using dictionary learning for classification; that is to simultaneously maximise the discriminability and sparse-representability power of the learned dictionaries. Upon this premise, we designed class-specific dictionaries incorporating three factors: discriminability, sparsity and classification error. We integrated these metrics into a unified cost function and adopted a new feature space, i.e., histogram of oriented gradients (HOG), to generate the dictionary atoms. The rationale of using HOG features for designing the dictionaries is their strength in describing fine details of crowded images. The results of applying the proposed method in the classification of Chinese handwritten numbers demonstrated enhanced classification performance $(\sim98\%)$ compared to state-of-the-art deep learning techniques (i.e., SqueezeNet, GoogLeNet and MobileNetV2), but with a fraction of parameters. Furthermore, combination of the HOG features with dictionary learning enhances the accuracy by $11\%$ compared to the case where only pixel domain data are used. These results were supported when the proposed method was applied to both Arabic and English handwritten number databases.