Multimodal Deep Neural Networks for Digitized Document Classification

1  Introduction

Each day an increasing amount of massive data is generated on the Internet due to the regular development of social media, the Internet of Things (IoT), and mobile computing. Similar to conventional data types, such a quantity of growing information demands a proper categorization to manage it most efficiently. Automatic document classification is becoming more crucial for trading companies, stores, and information retrieval as the volume of electronic text documents increases incredibly fast. Since it often satisfies the requirements of the present applications better, multi-label classification is noticeably preferable to single-label classification.

Vast amounts of generated digital data are available for data centers to use for quicker, more efficient, and automated processing in this digital age [1]. Property records examination, which generally deals with scanned images of distinct sorts, such as Certificates, Bank loans, Contracts, Payments, etc., is one of the sectors that provide ample opportunities for the application of Artificial Intelligence. The main challenge for automated administration and processing of the information hidden within these papers is classification. Properly handled property data directly influence the economic health of a nation. The fact that the materials are of diverse types, differing lengths, unstructured formats, and are closely connected in terms of the highlighted content [2] makes categorization challenging. The role of sophisticated exploratory data mining techniques, methodical exploration, and feature engineering has become crucial to develop a multi-class classifier and achieving an advanced level of accuracy in this assignment. Using the state-of-the-art embeddings methodology, which converts unprocessed text data into n-dimensional feature space, authors of some studies carefully examined property-related materials from six categories. They created a high-dimensional feature space using a linear kernel support vector machine (SVM) [3–5].

Modern approaches often include a variety of pre-processing techniques, such as feature selection, dimensionality reduction [6], and accurate document representations [7] to narrow the feature set while maintaining a high level of classification precision during the process. Nevertheless, this pre-processing technique has posed many severe problems, such as data loss, increased extra requirements analysis, task dependency or approval process, etc.

In artificial intelligence, neural nets with deep learning are now quite widespread, and it has been suggested that these networks surpass numerous state-of-the-art methodologies without the need for any parameter estimation. This is mainly fair when it comes to image analysis [8]. Still, it has also been proved that they perform better in Natural Language Processing (NLP), which includes tokenization, character segmentation, named entity identification, and semantic role labeling [9]. Nevertheless, whether the published work involves their use for categorizing multi-label documents has yet to be defined. Consequently, the application of neural networks to the problem of multi-label document categorization of university documents is the primary objective of this article. In the wake of the digital era, higher education institutions are grappling with an overwhelming amount of documentation that requires efficient and effective classification. Tackling this issue, the current study puts forward a deep learning-enabled framework for automated document classification. This proposal fills a pivotal niche in higher education, responding to the necessity for a structured and automated system to manage a burgeoning quantity of academic documentation.

The proposed method offers several distinct contributions to the field. Notably, it capitalizes on deep learning capabilities to enhance the document classification process, pushing the boundaries of existing methodologies. By taking advantage of state-of-the-art techniques, this study seeks to improve the accuracy and efficiency of document categorization.

The current work also addresses an understated aspect of automated classification: Interactivity. By fostering an interface that classifies and interacts intuitively with users, the proposed method augments user experience, a clear advantage over its predecessors. Evidence illustrating this enhancement will be provided in the ensuing sections.

Understanding the need for future adaptability and improvement, potential issues associated with the proposed method are also discussed in this paper. Future research directions are outlined to mitigate these issues, thereby ensuring the sustainability of this approach in the face of evolving requirements and technologies.

The following sections detail the construction and application of the proposed deep learning framework, its experimental validation, and how it compares to existing approaches. Furthermore, a comprehensive discussion will clarify its advantages, limitations, and potential areas for future enhancement, highlighting the study’s significance and contributions to automated document classification in higher education. The novelty of the proposed research is the following:

A novel multimodal deep learning model has been developed in this research, which innovatively integrates visual and textual data for document classification. This advancement supersedes previous unimodal techniques, exploiting the benefits of both data types and achieving higher classification performance.

A unique feature extraction method has been introduced in the proposed model, maintaining the inherent structure and semantics of digitized documents. This methodology improves the accuracy of classification.

The challenge of handling imbalanced datasets, a frequent issue in document classification, has been addressed through a new method formulated in our research. This technique enhances the model’s ability to classify under-represented classes, significantly boosting its versatility and applicability.

Furthermore, our study proposes an exhaustive evaluation framework for digitized document classification models. This approach considers the accuracy and interpretability of the model, promoting easier adoption in various practical applications.

The rest of this paper includes the following sections: Next section reviews the literature on categorizing digitized documents using deep learning methods. Section 3 deals with materials and methods, including the overall flowchart of the research, collected dataset, and proposed deep learning method for scanned document categorization. Following this, Section 4 provides obtained results, and at the end, the current research will be concluded, indicating the proposed method’s advantages and research findings.

2  Literature Review

The categorization of documents is often accomplished via the use of artificial intelligence techniques. These techniques often employ labeled data to train classification models, and subsequently, the classifications are applied to texts that have not been labeled yet. The Vector Space Model, which typically depicts each document with a vector of all word occurrences weighted by their Term Frequency-Inverse Document Frequency (TF-IDF), is one of the widespread methods used in most published articles.

There have been several successful applications of classification techniques, such as Bayesian classifiers, Maximum Entropy (ME), Support Vector Machines (SVMs), and other similar approaches [10]. However, the most significant challenge presented by this job is the large dimensionality of the feature space inside the VSM, which reduces the accuracy of the classifier.

In recent times, “deep” Neural Networks (NN) have shown their better efficiency in a variety of NLP problems, particularly POS tagging, character segmentation, representation-based identification, and semantic role labeling [11], all of which do not need any parametrization. Quite a few distinct architectures and learning methods were considered [12].

For instance, the researchers in [13] propose the utilization of two distinct convolutional neural networks (CNN) for tasks such as ontology construction, sentiment interpretation, and uni-label document categorization. Their networks comprise nine layers, with six being convolutional and the other three fully connected. Each layer has a distinct number of hidden units and a varied frame size. Applying both the English and the Chinese corpora, they demonstrate that the suggested strategy achieves much better results than the baseline methods (bag of words). Another fascinating piece of research [14] leverages pre-trained vectors from word2vec [15] in the first layer of their system. The authors determined that the suggested models perform better than the current state-of-the-art on four out of seven tasks, which include question categorization and sentiment analysis.

Traditional neural networks with numerous layers were also used for multi-label document categorization [16]. The authors have developed a unique error function and implemented it into a modified version of the classic backpropagation method for multi-label learning. The effectiveness of this method is examined using functional genomics and text classification. Within the subset of bottom-up methodologies, studies [17–19] highlight the efficacy of connected component analysis in tandem with the sliding-window technique. This synergy, coupled with the extraction of textural attributes, has exhibited notable success. These algorithms are relatively resistant to skew, although the processing they require may change depending on the selected metrics. According to reports, the accuracy of text detection ranges from 94% to 99%.

Problems in discriminating between many classes are often related to zone classifications. Haleem et al. [20] categorize eight representative items about the scanned texts. The experiment’s findings have shown that successfully carrying out this activity is challenging. The reported mistake rate achieved 2.1%, yet 72.7% of the logos and 31.4% of the tables needed to be corrected. Although the other study showed a relatively high mean accuracy of 98.45%, 84.64% of trademarks and 72.73% of ‘other’ items were incorrectly identified [21].

An examination of the relevant published material reveals that most of the methodologies in question use contrast enhancement methods, only apply to specific types of documents, and combine hand-crafted features with multi-tiered methodologies [22]. The second observation revealed the need for more techniques to identify many fascinating elements that may be found in scanned texts. The main reason was that the properties of things have a great deal of variation regarding how they seem in different contexts.

Convolutional neural networks that can cope with these challenges using a particular approach were developed to overcome the abovementioned limits. This different study path is targeted at producing efficient classifiers rather than paying attention to constructing well-created low-level characteristics [23]. The challenge of translating model parameters, such as visual attributes, like the intensity of pixels, to specified outputs, such as an object type of object, is our task, and this study proposes a possible solution in this kind of situation. Due to the hierarchical structure in perceptual representation, the deep learning technique contains sets of inputs, hidden, and outputs. A model of this kind integrates high-level abstract notions with lower-level, more fundamental characteristics. One of the primary benefits of using such a strategy is the flexibility of learning. It is essential to emphasize that learning is not wholly innovative since it derives from traditional machine learning methods. Iterative data is supplied into the input of a deep neural network. Afterward, the training process allows the network to calculate layer by layer to create output, which is then compared with the right answer. Backpropagation is a process that corrects the weights of each node in a neural network by moving the output error backward through the network. This helps to minimize overall error. This approach makes incremental improvements to the model while it is being computed. The convolution of a feature descriptor with input data is the most critical computational issue a CNN must solve. Therefore, the characteristics go from simple pixels to particular primitives such as lateral and vertical boundaries, circles, and color regions as they go up the hierarchy. On the other hand, CNN filters process all of the input vectors simultaneously, in contrast to traditional image filters, which only function on single-channel pictures. As a result of the multilayer filters’ interpretation, they provide a robust response in regions in which a particular feature is encountered. Table 1 presents a comparative analysis of contemporary research in the field of document classification.

3  Materials and Methods

This section contemplates every aspect of our approach to subject identification in depth. Additionally, it is advised to undertake a textual and visual component-based study of the original document to develop and deploy a unique semantic topic identification technique. The rest of this section is devoted to a thorough description of the system architecture, focusing on the fundamental traits each module of the suggested framework share. In addition, it also provides extensive details regarding the multimedia knowledge base and the ontological model, which was employed as a basis for the base. Next, a comprehensive explanation of the presented Topic Detection method is provided.

Fig. 1 demonstrates the flowchart of the proposed system for the document categorization problem. A dataset comprising 8139 images, categorized into seven distinct classes, was collected for this study. Following this, DOM Parser was applied to facilitate accessing and modifying XML documents’ style, structure, and content. Text documents were then preprocessed and added to the data repository. This process succeeded by collecting pertinent features required for training a categorization model. With the necessary features identified, a deep learning model was developed for training and testing. Consequently, categorized documents were distributed into distinct classes.

Figure 1: System architecture

This allows the taxonomy classification to construct a classification taxonomy beginning with a concept. Comparisons have been drawn between the suggested measure, method, and baselines; the findings are illustrated and discussed in experimental Section 4 of the report.

Fig. 2 illustrates the proposed deep neural network to tackle the document categorization problem. The network’s input constitutes a scanned document. This network comprises 11 layers, with the first layer functioning as a pooling layer. Subsequently, a Conv2D layer is utilized, and the resulting output is forwarded to the bottleneck layers. Within this network, there are five bottleneck layers. Following this, another pooling layer and a Conv2D layer are employed. The subsequent stage encompasses 128 neurons, while the penultimate layer comprises only seven layers, representing the document classes. The categorized class is ultimately determined using Maxpooling.

Figure 2: A multimodal classification system tailored for integrated text and image categorization. The training process seamlessly incorporates both linguistic and visual attributes

Fig. 3 offers a depiction of document samples drawn from the amassed dataset. Displayed in Fig. 3 are three distinct document types: A service letter, a diploma certificate, and a personnel document. This dataset comprises seven different types of university documents. With 8139 scanned documents spanning seven categories, the dataset occupies more than 4.7 Gb of storage.

Figure 3: Examples of document categories within the curated dataset

4  Experiment Results

This section presents the results derived from the conducted experiments, offering a comprehensive overview of the performance and efficacy of the multimodal deep learning model. Detailed analyses have been conducted on diverse datasets, allowing us to evaluate the robustness and versatility of our methodology. Our findings and a thorough comparison with other existing techniques offer valuable insights into our proposed model’s strengths and areas for improvement. The study mainly entails seven primary categories of university documents included in the cases handled by the STF. Below is a listing of these categories, with their original labels preserved.

4.1 Experimental Setup and Dataset

It should be noted that the legal cases include various papers, all of which were categorized under the heading “Miscellaneous.” In this context, an annotation tool was developed and subsequently employed by a group of four attorneys to manually categorize a collection of 8,139 documents. The percentage of papers that belong to each category is broken out graphically in Fig. 4, which may be seen below.

Figure 4: Categorization of document types within the dataset

It is standard practice to divide datasets into three sections to train and assess machine learning systems [31]. These divisions of the dataset are identified as the train, validation, and test subsets. Stratified splits are applied for each document class, thereby ensuring the inclusion of proportional class samples in each subset. The following ratios were employed, which are broken out further in Fig. 5:

Figure 5: Distribution of the accumulated data set based on the occupied disk space

•   70% is allocated for the training subset,

•   20% is designated for validation purposes, and

•   10% is reserved for the testing subset.

Fig. 4 presents the dispersion of the amassed dataset based on document quantity. The dataset comprises seven distinct categories of scanned materials, which are not balanced. Predominantly, the dataset is enriched by student-related directives, accounting for approximately 35% of the total scanned entities. Subsequently, the Journal of Accounting for advanced academic credentials, individual records, and manufacturing directives represent 20%, 19%, and 13% of the entire dataset, respectively. Notably, general orders, manufacturing directives, and correspondence related to services occupy the smaller segments of the dataset, contributing 6%, 4%, and 13%, respectively.

All the collected data composes of about 4.7 Gb of scanned pdf data. Fig. 5 illustrates the partition of these data by their volumes for each categorization. Diplomas, Personal documents, and higher education diplomas constitute 87.8% of all dataset volumes, with 42.6%, 19%, and 26%, respectively. The other four types of collected documents compose 12.2%.

Fig. 6 showcases the allocation of the training, validation, and test subsets utilized for training the advanced deep learning model. The amassed dataset was divided into three parts: Training, validation, and test sets, comprising 70%, 20%, and 10% of the total data, respectively. Thus, dividing the dataset into three parts allows us to get high accuracy and understand whether the proposed model is effective in practice.

Figure 6: Distribution of training, validation, and testing subsets across each document category

Table 2 summarizes Figs. 5 and 6, illustrating the dispersion of the amassed dataset based on the count of scanned visuals and the data volume for every distinct category. It allows us to understand the relation of the number of images to the volume quality of the document by type. Thus, the model’s training can commence following the dataset’s preparation.

4.2 Evaluation Parameters

This section aims to elucidate the evaluation parameters used to assess the proposed model and facilitate its comparison with other machine learning models. The following five indicators were selected as evaluation parameters: Accuracy, precision, recall, F-score, and area under the curve receiver operating characteristics (AUC-ROC). Eq. (1) stands for the formula of accuracy.

Accuracy=TP+TNP+N(1)

Eq. (2) shows the formula of precision.

Precision=TPTP+FP(2)

Eq. (2) indicates the formula of recall.

Recall=TPTP+FN(3)

Eq. (2) demonstrates the formula of the F-score.

F1=2×precision×recallprecision+recall(4)

4.3 Results

This section presents the evaluation results of the proposed deep learning model for classification of scanned documents. Subsequently, the proposed model’s model accuracy, validation accuracy, and confusion matrix are presented. Fig. 7 illustrates the accuracy of the model over the training of 100 epochs. The findings indicate that the suggested model exhibits commendable precision during both training and testing phases.

Figure 7: Accuracy of the model

Fig. 8 depicts the model’s loss across 100 epochs. The data suggests a swift decline in loss as the epochs increase. By the 80th epoch, both training and testing losses register below 0.3. This underscores the model’s suitability for validated instances and its potential for effectively addressing automated document classification challenges.

Figure 8: Model loss progression chart

Fig. 9 presents the confusion matrix corresponding to each categorized document type, serving as a performance assessment metric for the machine learning classification task that involves multi-class outputs. The table underneath contains the four possible permutations of anticipated and actual values. The results indicate that there is a minimum number of errors and confusion. In most cases, scanned documents are categorized correctly.

Figure 9: Confusion matrix

Table 3 contrasts the advanced deep learning model proposed with other established machine learning methodologies. The results claim that the proposed model gives the highest performance comparing the other methods in each evaluation parameter. The advanced model proposed yields performance metrics of 94.84% in accuracy, 94.79% in precision, 94.62% in recall, 94.43% in F-score, and 94.07% in AUC-ROC. These figures indicate the model’s real-world applicability, and the provided dataset is apt for training both machine learning and deep learning frameworks in addressing document classification challenges.

Fig. 10 juxtaposes the advanced deep learning model proposed with five conventional machine learning methods, specifically: XGBoost, multilayer perceptron, random forest, support vector machines, and decision tree, all targeting document classification tasks. Multiple evaluation metrics, including accuracy, precision, recall, F-score, ROC-AUC, and threshold, are employed for this comparison. The results suggest that the advanced deep learning model consistently outperforms across all these metrics. As a result, this deep model proves to be effective for classifying academic institutional documents.

Figure 10: Derived outcomes juxtaposed with traditional machine learning techniques

Table 4 compares the proposed method with state-of-the-art studies. Different studies have been developed for scanned document categorization of electronic health records, magnetic resonance images, and handwritten documents. The proposed Conv2D model categorizes educational documents of seven types with 94.84% accuracy.

In the pursuit of empirical validation, the proposed computational model was rigorously evaluated against the benchmark ICDAR 2017 competition dataset [40]. The comparative analysis is systematically presented in Table 5, where the performance metrics of the model are juxtaposed with those derived from the aforementioned dataset. This analytical juxtaposition corroborates the robustness of the proposed model, manifesting its competence in transcending dataset-specific constraints. The results unequivocally indicate that the model does not only retain its efficacy across diverse datasets but also achieves commendable evaluation scores in the domain of scanned document categorization, thus reinforcing the potential for wide applicability and the generalizability of its algorithmic constructs.

5  Conclusion

In the current era, where the challenge is to manage an overwhelming amount of data, there exists a pressing need for AI technologies capable of classifying documents. During the extraction of features, such tools need to be able to swiftly and efficiently offer access to the searched data. To address this issue, our study proposes a method for identifying associated metadata. In contrast to other works, it incorporates verbal and visual elements into its structure. In addition, the study uses statistics and semantics to recognize the topics of multimedia online publications. In particular, using semantics enhances the value of our findings by elevating them to a higher level and immediately addressing a problem already well-known as the semantic gap. The proposed strategy yields high performance in various circumstances since it can assist in the organization of a document collection and describe the primary topic of a document through semantic multimedia analysis among concepts. Both of these capabilities contribute to the achievement of high-level performance. Consequently, an array of tests was executed to evaluate the effectiveness of various topic identification tasks, with the outcomes discussed in detail. In this regard, these results substantiate the superiority of the presented method for text topic detection compared to state-of-the-art methods concerning the subject at hand, such as LSA and LDA. Regarding visual topic identification, numerous types of descriptors have been rigorously tested. The most promising findings were selected, particularly those based on the characteristics derived from the activation layer of the deep neural network. Moreover, it was demonstrated that combining the strategy for textual topic identification with that of visual topic identification can feasibly enhance the overall task by leveraging the most successful elements of both strategies. Moreover, it can adjust the suggested architecture to implement multiple models for topic identification since the system is modular and can be reused. This task can be accomplished by adding new modules or updating the already present ones. The system has successfully passed all tests using a representative sample of online documents. However, the system’s architecture permits the use of various libraries containing multimedia materials.

As part of our ongoing research, our next objective is to expand the evaluation of the proposed framework using a variety of document sets shortly. In addition, considering how difficult it is to locate data sets for websites with both text and multimedia elements, one of the following tasks will be constructing such a dataset while simultaneously using our knowledge base to identify discussion topics. Furthermore, several methods of subject identification will be investigated in the following days. Addressing the problem of multilingualism to enhance the performance of textual topic recognition is an intriguing area of study that needs to be pursued. In particular, the computational efficiency of the textual topic recognition method will be the primary focus of our attention. To advance the current state of the art in this field, there are plans to research innovative approaches and algorithms for visual topic recognition.

Acknowledgement: Not applicable.

Funding Statement: The authors received no specific funding for this study.

Author Contributions: Conceptualization, A.B.; methodology, A.B., A.S., and B.M.; software, A.B.; data curation, A.B.; writing—original draft preparation, A.B., and A.S.; writing—review and editing, A.B., N.G., R.N., N.Z.; supervision, A.Z.

Availability of Data and Materials: The data used in this paper can be requested from the corresponding author upon request.

Conflicts of Interest: The authors declare that they have no conflicts of interest to report regarding the present study.

References

1. H. Jain, S. Joshi, G. Gupta, and N. Khanna, “Passive classification of source printer using text-line-level geometric distortion signatures from scanned images of printed documents,” Multimed. Tools Appl., vol. 79, no. 11, pp. 7377–7400, 2020. [Google Scholar]

2. D. Sultan et al., “A review of machine learning techniques in cyberbullying detection,” Comput. Mater. Contin., vol. 74, no. 3, pp. 5625–5640, 2023. [Google Scholar]

3. G. Wiedemann and G. Heyer, “Multi-modal page stream segmentation with convolutional neural networks,” Lang. Resour. Eval., vol. 55, no. 1, pp. 127–150, 2021. [Google Scholar]

4. S. Chattopadhyay et al., “MTRRE-Net: A deep learning model for detection of breast cancer from histopathological images,” Comput. Biol. Med., vol. 150, pp. 106155, 2022 [Google Scholar] [PubMed]

5. A. Basu, K. Sheikh, E. Cuevas, and R. Sarkar, “COVID-19 detection from CT scans using a two-stage framework,” Expert Syst. Appl., vol. 193, pp. 1–14, 2022. [Google Scholar]

6. B. Omarov, S. Narynov, Z. Zhumanov, A. Gumar, and M. Khassanova, “A skeleton-based approach for campus violence detection,” Comput. Mater. Contin., vol. 72, no. 1, pp. 315–331, 2022. [Google Scholar]

7. R. das Neves, E. Lima, B. Bezerra, C. Zanchettin, and A. Toselli, “HU-PageScan: A fully convolutional neural network for document page crop,” IET Image Process., vol. 14, no. 15, pp. 3890–3898, 2020. [Google Scholar]

8. E. Chaowicharat and N. Dejdumrong, “A step toward an automatic handwritten homework grading system for mathematics,” Inf. Technol. Control, vol. 52, no. 1, pp. 169–184, 2020. [Google Scholar]

9. I. Nasir et al., “Pearson correlation-based feature selection for document classification using balanced training,” Sens., vol. 20, no. 23, pp. 1–18, 2020. [Google Scholar]

10. W. Zaaboub, L. Tlig, M. Sayadi, and B. Solaiman, “Neural network-based system for automatic passport stamp classification,” Inf. Technol. Control, vol. 49, no. 4, pp. 583–607, 2020. [Google Scholar]

11. A. Singh et al., “Joint encryption and compression-based watermarking technique for security of digital documents,” ACM Trans. Internet Technol. (TOIT), vol. 21, no. 1, pp. 1–20, 2021. [Google Scholar]

12. J. Latham et al., “Effect of scan pattern on complete-arch scans with 4 digital scanners,” J. Prosthet. Dent., vol. 123, no. 1, pp. 85–95, 2020 [Google Scholar] [PubMed]

13. A. Nurunnabi, F. Teferle, J. Li, R. Lindenbergh, and A. Hunegnaw, “An efficient deep learning approach for ground point filtering in aerial laser scanning point clouds, The Inter. Archives Photogramm. Remote Sens. Spatial Inf. Sci., vol. 43, pp. 31–38, 2021. [Google Scholar]

14. M. Doneus, G. Mandlburger, and N. Doneus, “Archaeological ground point filtering of airborne laser scan derived point-clouds in a difficult mediterranean environment,” J. Comput. Appl. Archaeol., vol. 3, no. 1, pp. 92–108, 2020. [Google Scholar]

15. G. Malaperdas, “Digitization in archival material conservation processes,” Eur. J. Eng. Technol. Res., vol. 6, no. 4, pp. 30–32, 2021. [Google Scholar]

16. D. Zhang et al., “A multi-label classification method using a hierarchical and transparent representation for paper-reviewer recommendation,” ACM Trans. Inf. Syst., vol. 38, no. 1, pp. 1–20, 2020. [Google Scholar]

17. T. Hegghammer, “OCR with tesseract, amazon textract, and google document AI: A benchmarking experiment,” J. Comput. Soc. Sci., vol. 5, no. 1, pp. 861–882, 2022. [Google Scholar]

18. M. Revilla-León, M. Sadeghpour, and M. Özcan, “An update on applications of 3D printing technologies used for processing polymers used in implant dentistry,” Odontology, vol. 108, no. 3, pp. 331–338, 2020. [Google Scholar]

19. C. Mangano, F. Luongo, M. Migliario, C. Mortellaro, and F. Mangano, “Combining intraoral scans, cone beam computed tomography and face scans: The virtual patient,” J. Craniofac. Surg., vol. 29, no. 8, pp. 2241–2246, 2018 [Google Scholar] [PubMed]

20. A. Haleem and M. Javaid, “3D scanning applications in medical field: A literature-based review,” Clin. Epidemiol. Glob. Health, vol. 7, no. 2, pp. 199–210, 2019. [Google Scholar]

21. W. Hassan, Y. Yusoff, and N. Mardi, “Comparison of reconstructed rapid prototyping models produced by 3-dimensional printing and conventional stone models with different degrees of crowding,” Am J. Orthod. Dentofac., vol. 151, no. 1, pp. 209–218, 2017. [Google Scholar]

22. K. Adam, A. Baig, S. Al-Maadeed, A. Bouridane, and S. El-Menshawy, “KERTAS: Dataset for automatic dating of ancient Arabic manuscripts,” Int. J. Doc. Anal. Recognit, vol. 21, no. 4, pp. 283–290, 2018. [Google Scholar]

23. X. Liu et al., “Automatic organ segmentation for CT scans based on super-pixel and convolutional neural networks,” J. Digit. Imaging, vol. 31, no. 5, pp. 748–760, 2018 [Google Scholar] [PubMed]

24. C. Ben Rabah, G. Coatrieux, and R. Abdelfattah, “Automatic source scanner identification using 1D convolutional neural network,” Multimed. Tools Appl., vol. 81, no. 16, pp. 22789–22806, 2022. [Google Scholar]

25. W. Qin, R. Elanwar, and M. Betke, “Text and metadata extraction from scanned arabic documents using support vector machines,” J. Inf. Sci., vol. 48, no. 2, pp. 268–279, 2022. [Google Scholar]

26. A. Naïve and J. Barbosa, “Efficient accreditation document classification using naïve bayes classifier,” Indian J. Sci. Technol., vol. 15, no. 1, pp. 9–18, 2022. [Google Scholar]

27. A. Kumar et al., “Closing the loop: Automatically identifying abnormal imaging results in scanned documents,” J. Am. Med. Inf. Assoc., vol. 29, no. 5, pp. 831–840, 2022. [Google Scholar]

28. F. Rashid, S. Gargaare, A. Aden, and A. Abdi, “Machine learning algorithms for document classification: Comparative analysis,” Int. J. Adv. Comput. Sci. Appl., vol. 13, no. 4, pp. 260–265, 2022. [Google Scholar]

29. A. Guha, A. Alahmadi, D. Samanta, M. Khan, and A. Alahmadi, “A multi-modal approach to digital document stream segmentation for title insurance domain,” IEEE Access, vol. 10, no. 1, pp. 11341–11353, 2022. [Google Scholar]

30. S. Shreesha and H. Anita, “Classification of north and south Indian handwritten scripts using gabor wavelet features,” Indian J. Sci. Technol., vol. 15, no. 16, pp. 712–717, 2022. [Google Scholar]

31. B. Omarov et al., “Artificial intelligence in medicine: Real time electronic stethoscope for heart diseases detection,” Comput. Mater. Contin., vol. 70, no. 2, pp. 2815–2833, 2022. [Google Scholar]

32. E. Hsu, I. Malagaris, Y. Kuo, R. Sultana, and K. Roberts, “Deep learning-based NLP data pipeline for EHR-scanned document information extraction,” JAMIA Open, vol. 5, no. 2, pp. pp 1–12, 2022. [Google Scholar]

33. G. BinMakhashen and S. Mahmoud, “Historical document layout analysis using anisotropic diffusion and geometric features,” Int. J. Digit. Libr., vol. 21, no. 3, pp. 329–342, 2020. [Google Scholar]

34. S. Darwish and H. ELgohary, “Building an expert system for printer forensics: A new printer identification model based on niching genetic algorithm,” Expert Syst., vol. 38, no. 2, pp. 1–14, 2021. [Google Scholar]

35. K. Aderghal, K. Afdel, J. Benois-Pineau, and G. Catheline, “Improving Alzheimer’s stage categorization with convolutional neural network using transfer learning and different magnetic resonance imaging modalities,” Heliyon, vol. 6, no. 12, pp. 1–13, 2020. [Google Scholar]

36. K. Han et al., “An expert-in-the-loop method for domain-specific document categorization based on small training data,” J. Assoc Inf. Sci. Technol., vol. 74, no. 6, pp. 669–684, 2023. [Google Scholar]

37. S. Khaleefah, S. Mostafa, A. Mustapha, and M. Nasrudin, “The ideal effect of gabor filters and uniform local binary pattern combinations on deformed scanned paper images,” J. King Saud Univ.-Comput. Inf. Sci., vol. 33, no. 10, pp. 1219–1230, 2021. [Google Scholar]

38. A. Jadli, M. Hain, and A. Hasbaoui, “An improved document image classification using deep transfer learning and feature reduction,” Int. J., vol. 10, no. 2, pp. 549–557, 2021. [Google Scholar]

39. G. Nagy, “Twenty years of document image analysis in PAMI,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 22, no. 1, pp. 38–62, 2000. [Google Scholar]

40. R. Gomez et al., “ICDAR2017 robust reading challenge on COCO-Text,” in Proc. 2017 14th IAPR Int. Conf. Document Anal. Recognit. (ICDAR), Kyoto, Japan, 2018, pp. 1435–1443. [Google Scholar]

41. K. Hashmi, A. Pagani, M. Liwicki, D. Stricker, and M. Afzal, “Cascade network with deformable composite backbone for formula detection in scanned document images,” Appl. Sci., vol. 11, no. 16, pp. 7610, 2021. [Google Scholar]

42. J. Younas et al., “Fi-Fo detector: Figure and formula detection using deformable networks,” Appl. Sci., vol. 10, no. 18, pp. 6460, 2020. [Google Scholar]

43. L. Gao, X. Yi, Z. Jiang, L. Hao, and Z. Tang, “ICDAR2017 competition on page object detection,” in Proc. 2017 14th IAPR Int. Conf. Document Anal. Recognit. (ICDAR), Kyoto, Japan, 2017, pp. 1417–1422. [Google Scholar]

44. X. Li, F. Yin, and C. Liu, “Page object detection from pdf document images by deep structured prediction and supervised clustering,” in Proc. 2018 24th Int. Conf. Pattern Recognit. (ICPR), Beijing, China, Aug. 20–24, 2018, pp. 3627–3632. [Google Scholar]

45. N. Ranjan, “Document classification using LSTM neural network,” J. Data Min. Manag., vol. 2, no. 2, pp. 1–9, 2017. [Google Scholar]