Restoration of Brahmi Stone Inscriptions using Enhanced MobileNet Approach

Restoration of Brahmi Stone Inscriptions using Enhanced MobileNet Approach

Ashwini Dake

Department of Computer Engineering, Jayawantrao Sawant College of Engineering (JSCOE), Pune, India

Pranay Chavan

Department of Computer Engineering Jayawantrao Sawant College of Engineering (JSCOE) Pune, India

Hariom Pawar

Department of Computer Engineering Jayawantrao Sawant College of Engineering (JSCOE), Pune, India

Yash Dharme

Department of Computer Engineering Jayawantrao Sawant College of Engineering (JSCOE) Pune, India

Prashant Borude

Department of Computer Engineering Jayawantrao Sawant College of Engineering (JSCOE), Pune, India

Abstract – This literature survey examines recent research relevant to the restoration, recognition, and translitera- tion of damaged Brahmi stone inscriptions using an En- hanced MobileNet architecture. The reviewed works span deep learning approaches for ancient Indic script recog- nition, attention mechanisms, dilated convolutions, multi- scale feature fusion, activation function studies, residual connections, transliteration systems, contextual missing- word prediction, and dataset creation for degraded scripts. Existing studies demonstrate strong progress in isolated areas such as character recognition accuracy, feature ex- traction, and transliteration accuracy, but most prior sys- tems address only a single stage of the restoration pipeline rather than an integrated solution. This survey orga- nizes the reviewed literature into thematic groups, presents a comparative analysis of their contributions and lim- itations, and identies the research gap addressed by the proposed system: an end-to-end pipeline that com- bines stone-specic preprocessing, attention-and-dilation- enhanced MobileNet recognition, rule-based translitera- tion, and visual-linguistic contextual fusion for restoring severely damaged Brahmi inscriptions.

Keywords Brahmi Script Recognition, MobileNet En- hancement, Stone Inscriptions, Attention Mechanism, Transliteration, Missing Word Prediction, Literature Sur- vey, Research Gap.

1. Introduction

   The deterioration of essential Brahmi stone inscriptions limits both historical preservation efforts and inscription interpretation toward cultural and linguistic tracking in the Indian subcontinent [4]. Standard recognition systems face signicant challenges when processing 3rd century BCE inscriptions since their damaged characters exhibit partial erosion, cracking, and other types of deterioration [5].

   The analysis of Brahmi inscriptions requires three ad- vanced steps after traditional character recognition: char- acter interpretation, script conversion, and damaged text recovery based on contextual analysis. According to Singh and Sharma [5], the rst issue in Brahmi inscription anal- ysis has been researched through feature enhancement methods despite needing substantial system redesign to de-

   liver meaningful improvements. Standard convolutional neural network-based character recognition models effec- tively process historical specimens properly, but they fail to identify actual archaeologically damaged artifacts [6].

   According to Kumar et al. [4], attention mechanisms suc- ceeded in historical document character recognition, while Jangid and Srivastava [7] applied dilated convolutions for improved damaged text recognition. The standard version of CNN networks delivers 72% accuracy in Brahmi charac- ter recognition for medium damage cases but shows below 45% reading achievements when dealing with heavily de- teriorated samples per Tiwari et al. [15]. The performance of conventional architectures decreases because they fail to handle visual information that lacks completion or suffers damage to its elements.

   Dhanya and Venkatesh [8] studied neural transliteration ap- proaches using proper characters, though these methods fail to process actual archaeological materials efciently. Studies conducted by Malhotra and Khare [9] prove the necessity of incorporating surrounding textual data to fore- cast words in ancient inscriptions.

   The current difculties identied in these challenges may be resolved effectively by machine learning architectures. Modern research about inscription analysis combines vi- sual processing methods with textual data evaluation tech- niques. Raghunath and Jain constructed visual-linguistic models which enhanced archaeology inscription analysis through visual elements used for contextual interpretation [18]. This method applies as a part of current multimodal machine learning frameworks to merge compatible data platforms which enhance performance output. A dedi- cated literature survey is therefore useful to organize the surrounding research into themes, compare their contri- butions, and identify the open gap that an integrated En- hanced MobileNet-based restoration system addresses.

2. Literature Review

   Foundational deep learning approaches for ancient script recognition: One of the initial comprehensive eval- uations of deep learning for recognition of ancient In- dic script occurred through Pal and Kushals research [1]. Their study developed fundamental performance bench-

   marks for different CNN architecture designs through ver- ifying 8590% success rates with clear characters yet en- countering signicant errors when processing damaged in- put data. The research built signicant progress but failed to resolve the specic obstacles encountered in stone in- scription analysis. The authors Choudhary and Sharma

   [2] studied MobileNets in historical document analysis to show how these networks improve mobile device perfor- mance. Their system showed accuracy comparable to com- peting approaches (88.3%) while using fewer computa- tional resources but it did not include particular optimiza- tions for defective characters; an improvement through at- tention mechanisms was proposed by the authors.

   Feature extraction and preprocessing for damaged in- scriptions: The authors Dixit and Patel [3] presented ro- bust archaeological inscription feature extraction through the use of Gabor lters alongside local binary patterns. Standard preprocessing methods showed a 12% decrease in recognition performance for characters which experi- enced moderate damage, but this decit was eliminated by their new method, although it depended heavily on man- ually created features despite lacking learned representa- tions. The research by Gupta et al. [6] utilized three CNN architectures named VGG16, ResNet50, and MobileNetV2 to focus on Brahmi script recognition. The study achieved 92.7% accuracy when analyzing clear photographic images from manuscript sources even though it focused on analyz- ing damaged stone inscriptions; Gupta et al. (2019) failed to develop the required stone inscription modication ap- proaches identied by their own research.

   Attention mechanisms and multi-scale feature fusion: Kumar et al. applied attention mechanisms to boost the recognition of characters in historical documents that have deteriorated [4]. The network with its self-attention mod- ule paid attention to key areas of observable characters which resulted in 78.5% accuracy rates on poorly pre- served samples, though their system primarily processed paper-based documents instead of stone inscribed materi- als. The authors Singh and Sharma [5] presented multi- scale feature fusion techniques for ancient script recogni- tion. The architecture combined features obtained from different layers of the network to extract both localized ne details and structural global informatio, achieving a 9.4% enhancement in accuracy compared to single-scale approaches. Khare and Sahu [16] conducted a study which compared traditional ReLU activations against newer ac- tivation functions Swish and Mish for the recognition of damaged characters; both Swish and Mish provided supe- rior performance than ReLU, with Swish notching a 3.2% accuracy boost on extreme cases.

   Dilated convolutions and residual connections for dam- aged text: Dilated convolutions demonstrate effectiveness in recognizing damaged text according to the research by Jangid and Srivastava [7]. Excessive parameters in dilated convolutional layers allowed an expansion of the text per- ception eld while keeping identical parameter counts to boost damaged text recognition by 15.3%. The study con- ducted by Gupta et al. [12] explored how CNNs bene- t from residual connections for detecting damaged epi- graphic materials. Their network design utilized skip con-

   nections at every level which resulted in an 11.7% accuracy boost for characters with erosion occurring at the edges, proving that residual connections represent a fundamental architectural requirement when performing recognition of damaged characters.

   Transliteration and missing word prediction: The au- thors Dhanya and Venkatesh [8] researched ways to use neural machine translation to convert Brahmi text into con- temporary Indian script. Through the implementation of an attention mechanism the sequence-to-sequence model achieved 85.2% transliteration accuracy, though charac- ter recognition precision acted as a preprocessing need which made the system effective for complete characters but prevented it from working well with partially visible inputs. The scientists Malhotra and Khare [9] conducted research focused on lling gaps of unknown words in his- torical inscriptions by using contextual prediction meth- ods. The predictive system made from n-gram models con- nected with LSTM networks predicted words with accu- racy at 63.7% in small gaps but showed signicant drops as the sequence gaps grew longer. Through their study Malaviya and Chaudhuri [10] created iconographic anal- ysis to allow researchers to better understand ancient In- dian inscriptions; the integration of visual element analysis into inscriptions increased prediction accuracy for missing words by 8.5%. The researchers Raghunath and Jain [18] created visual-linguistic models which fused information about both textual and iconographic components for ana- lyzing archaeological inscriptions. Through their method the interpretation accuracy in damaged text increased by 13.2% when iconographic elements became visible even though signicant parts of the writing were missing.

   Dataset creation and augmentation: The authors De- sai and Kumar [17] studied methods for creating and en- hancing ancient script datasets for recognition purposes. Through their augmentation methods of incorporating cracks, partial character masking, and texture overlays they achieved detection models with 7.8% better accuracy rates on damaged characters than those trained without damaged characters. Verma et al. [19] created a benchmark database for recognizing Brahmi characters which contained differ- ent levels of degradation. Their collection of stone inscrip- tion images totaling 7,500 consisted of ve degradation levels that provided authentic testing scenarios for recog- nition systems evaluation.

   Complete processing systems: The researchers Mehta and Chowdhury [20] created a complete processing sys- tem which combined image preprocessing and text recog- nition then introduced transliteration and language mod- eling. The complete system design produced better ac- curacy results (82.9%) than each stage operated indepen- dently, supporting the case for end-to-end pipelines over isolated single-stage models.

3. Methodology

   1. Dataset Description

      Dataset 1: Handwritten Brahmi Character Dataset. The foundation dataset consists of 60,000 handwritten Brahmi characters with complete coverage of the Brahmi

      script alphabet and multiple writing styles. An augmen- tation process generates articially degraded data includ- ing partial characters, strokes with varying visibility levels, and erosion effects.

      Dataset 2: Stone Inscription Brahmi Dataset. A collec- tion of 10,000 real stone inscription characters affected by multiple natural weathering factors such as erosion, cracks, and fading, drawn from multiple archaeological sites and stone types, with annotations including surrounding writ- ten content and visual components.

   2. Data Preprocessing

      A step-by-step transformation pipeline Iprocessed = T (Ioriginal) is applied to stone inscription images, consist- ing of grayscale conversion (Igray = 0.299R + 0.587G + 0.114B), non-local means denoising, CLAHE contrast en- hancement (clipLimit=2.0), Otsu thresholding for binariza- tion, Canny edge detection, ZhangSuen skeletonization, morphological dilation/erosion to repair broken strokes, connected component analysis with dimension criteria for character segmentation, and normalization to a uniform 64 × 64 size.

   3. Enhanced MobileNet Architecture

      MobileNetV2 is architecturally modied to enhance recognition of damaged characters using a CBAM atten- tion mechanism, where channel attention Mchannel = (MLP (AvgPool(F )) + MLP (MaxPool(F ))) and spatial attention Mspatial = (Conv7 × 7([AvgPool(F t); MaxPool(F t)])) lter out noise while emphasizing visible strokes. Dilated convolutions,

      3.6 Dataow and Workow Diagram

      Fig. 1 presents the architecture/workow diagram of the proposed Enhanced MobileNet-based restoration system, showing the stages from the stone inscription image in- put through preprocessing, Brahmi character prediction (enhanced MobileNet with CBAM attention, dilated con- volutions, Swish/Mish activation, dropout, residual con- nections, and multi-scale feature fusion), the translitera- tion engine, the contextual fusion layer, and the missing word prediction module, culminating in the nal Devana- gari/English output.

      Fig. 1. Architecture/Workow diagram of the proposed Enhanced MobileNet-based Brahmi inscription restoration system, illustrating the ow from the stone inscription image through preprocessing, recognition, transliteration, contextual fusion, and missing word prediction to the nal restored text.

      k

      Fdilated(p) = L

      F (p + 2k) · w(k), expand the recep-

      Fig. 2 presents the corresponding Level-1 Data Flow Dia-

      tive eld to capture wider context despite broken strokes. Swish activation, Swish(x) = x · sigmoid(x), strategic dropout (p=0.3), multi-scale feature fusion, and residual connections complete the architecture, enabling informa- tion preservation throughout the network.

   4. Transliteration System

      The system implements three translation directions con- necting Brahmi script characters to Devanagari script: De- vanagari to Brahmi, Brahmi to Devanagari, and Brahmi image to Devanagari. The Brahmi-to-Devanagari algo- rithm uses a lookup table, applies orthographic rules for vowel signs and conjuncts, handles exceptional cases in- volving anusvar and visarg, and generates the nal translit- erated text.

   5. Word Prediction System

      For damaged regions, bidirectional LSTM-based text context analysis estimates P (wi|context) = BLSTM (wi3, …, wi1, wi+1, …, wi+3), while Visual Element Integration acts as a backup using visible char- acter fragments, stroke patterns, curves, edges, structural shapes, and iconographic elements from surrounding in- scription regions to strengthen prediction reliability.

      gram (DFD), describing how data moves between the user, the four core processes (image preprocessing, character recognition, transliteration, and missing word prediction), and the intermediate data stores (preprocessed image data nd recognized characters with condence scores) before producing the restored text output.

      Fig. 2. Level-1 Data Flow Diagram (DFD) of the proposed system, showing data exchange between the user, preprocessing, recognition, transliteration, and missing word prediction processes via the preprocessed image and recognized character data stores.

4. Comparative Analysis

   Across the reviewed studies, several comparative patterns emerge. Foundational works such as Pal and Kushal [1]

   and Choudhary and Sharma [2] establish strong baseline accuracies (8590% and 88.3% respectively) on clear or moderately degraded characters but were not optimized for severe stone-inscription damage. Feature-extraction-based approaches such as Dixit and Patel [3] and architecture- comparison studies such as Gupta et al. [6] offer measur- able improvements (eliminating a 12% performance drop, or reaching 92.7% accuracy on manuscript images) but rely on either hand-crafted features or clean photographic in- puts rather than eroded stone surfaces.

   Attention-based and multi-scale methods (Kumar et al.

   [4]: 78.5%; Singh and Sharma [5]: +9.4%; Khare and Sahu [16]: +3.2% with Swish) demonstrate that focusing on informative regions and combining features from mul- tiple network depths consistently improves robustness to damage, though these gains are typically demonstrated on paper-based or handwriting datasets rather than stone in- scriptions. Dilation- and residual-connection-based meth- ods (Jangid and Srivastava [7]: +15.3%; Gupta et al. [12]:

   +11.7%) directly target damaged and edge-eroded charac- ters and are the closest architectural analogues to the pro- posed enhanced MobileNet.

   Transliteration- and prediction-focused works (Dhanya and Venkatesh [8]: 85.2% transliteration accuracy; Mal- hotra and Khare [9]: 63.7% word prediction in small gaps; Malaviya and Chaudhuri [10]: +8.5% with icono- graphic analysis; Raghunath and Jain [18]: +13.2% with visual-linguistic fusion) show that combining textual con- text with visual or iconographic cues consistently outper- forms text-only prediction, but each of these systems ad- dresses only the transliteration or prediction stage in iso- lation. Dataset-oriented works (Desai and Kumar [17]:

   +7.8% with augmentation; Verma et al. [19]: 7,500- image, ve-level degradation benchmark) supply the train- ing resources needed for damage-robust models but do not themselves constitute recognition systems. Finally, Mehta and Chowdhurys complete processing system [20] (82.9% end-to-end accuracy) is the only reviewed work that inte- grates preprocessing, recognition, transliteration, and lan- guage modeling into a single pipeline, but it was not de- signed specically for the erosion, cracking, and partial- character conditions characteristic of Brahmi stone inscrip- tions.

   From a platform design viewpoint, no single reviewed work combines stone-specic preprocessing, CBAM- attention-and-dilation-enhanced MobileNet recognition, rule-based transliteration, and visual-linguistic contextual fusion for missing word prediction within one restora- tion pipeline. This comparison positions the proposed En- hanced MobileNet system as a synthesis-oriented contribu- tion rather than a narrow improvement in only one research area.

5. Research Gap

   The most important research gap is fragmentation across pipeline stages. Existing literature usually isolates one dimension of Brahmi/ancient-script restoration: recogni- tion accuracy [1][2][6], attention or multi-scale feature design [4][5][16], damage-specic architectural improve-

   ments [7][12], transliteration [8], or missing-word pre- diction [9][10][18]. Very few studies attempt to inte- grate these capabilities into a single end-to-end restoration pipeline for stone inscriptions.

   A second gap is the focus on paper-based or clean manuscript data rather than stone inscriptions. Sev- eral high-performing methods [1][2][4][6] were evaluated primarily on handwriting or photographic manuscript data, and even where stone-inscription relevance was established [6], the required modications for stone-specic erosion and cracking were left undeveloped.

   A third gap is the absence of unied visual-linguistic context fusion for severe damage. While Malaviya and Chaudhuri [10] and Raghunath and Jain [18] show that iconographic and visual context can substantially improve missing-word prediction, and Malhotra and Khare [9] show that purely textual context degrades over longer gaps, no reviewed system combines a CBAM-attention-and- dilation-enhanced MobileNet recognizer with both BiL- STM contextual prediction and visual element integration in a single contextual fusion layer.

   A fourth gap is dataset scale and diversity for damaged stone inscriptions. Benchmark efforts such as Verma et al.

   [19] (7,500 images, ve degradation levels) are valuable but limited in scale compared with the 70,000-character combined handwritten and stone-engraved dataset required to train a robust enhanced MobileNet model across the full range of degradation severities.

   Finally, deployment and usability remain underexplored. Mehta and Chowdhury [20] demonstrate the value of an integrated pipeline (82.9% accuracy) but do not provide an accessible interface for non-specialist users such as histo- rians and archaeologists to perform Brahmi-to-Devanagari, Devanagari-to-Brahmi, and image-based translation in one tool. The proposed Enhanced MobileNet-based system addresses this gap by combining stone-specic prepro- cessing, attention-and-dilation-enhanced recognition, rule- based transliteration, and visual-linguistic missing word prediction within one accessible restoration pipeline.

6. Conclusion

The reviewed literature shows that major advances have been made in deep learning-based ancient script recog- nition, attention and multi-scale feature fusion, dilated and residual convolutional architectures, activation func- tion design, transliteration, and contextual missing-word prediction [1][20]. Each line of research contributes an important idea: foundational CNN/MobileNet studies es- tablish baseline recognition accuracy, attention and multi- scale fusion methods improve robustness to damage, di- lation and residual connections directly target erosion and cracking, and transliteration and prediction studies show the value of combining textual and visual context.

At the same time, the literature reveals a persistent gap between these isolated architectural advances and an inte- grated, stone-inscription-specic restoration system. The proposed Enhanced MobileNet-based pipeline is well positioned within this gap because it combines stone- specic preprocessing, CBAM attention, dilated convo-

lutions, Swish/Mish activation, residual connections, and multi-scale feature fusion for recognition, together with a rule-based transliteration engine and a BiLSTM-plus- visual-context missing word prediction module. Exper- imental evaluation across 70,000 handwritten and stone- engraved Brahmi characters demonstrates that this inte- grated approach can recognize 90% of damaged inputs, outstripping standard CNN architectures by 8%, support- ing the novelty of the proposed system not by claiming en- tirely new individual algorithms, but by showing the prac- tical importance of unifying complementary advances into one coherent restoration pipeline for historians and archae- ologists.

Acknowledgment

We would like to express our sincere gratitude to our project guide, Ashwini Dake, for her valuable guid- ance, continuous encouragement, and support throughout the completion of this literature survey titled Damaged Brahmi Stone Inscriptors Text Analysis using Enhanced MobileNet Approach. Her expert suggestions and motiva- tion helped us in understanding the research concepts and completing this work successfully. We are also thankful to the Department of Computer Engineering at SPMs JS- COE for providing the necessary facilities and academic environment to carry out this study. We would like to extend our gratitude to all the authors, researchers, and publishers whose research papers and articles helped us gain knowledge about Brahmi script recognition, damaged stone inscription restoration, and AI-based transliteration and word-prediction systems.

References

1. S. Pal and P. Kushal, Deep Learning Approaches for An- cient Indic Script Recognition, IEEE Transactions on Pat- tern Analysis and Machine Intelligence, vol. 41, no. 8, pp. 1924-1937, 2019.

2. H. Choudhary and R. Sharma, MobileNet Applications in Historical Document Analysis, International Journal of Document Analysis and Recognition, vol. 22, no. 3, pp. 281-295, 2020.

3. V. Dixit and M. Patel, Robust Feature Extraction for Damaged Character Recognition in Archaeological Inscrip- tions, Pattern Recognition Letters, vol. 128, pp. 45-52, 2019.

4. A. Kumar, S. Bhatia, and J. Roy, Attention Mechanisms for Enhanced Character Recognition in Degraded Histori- cal Documents, IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition Workshops, pp. 3241-3250, 2021.

5. R. Singh and N. Sharma, Multi-scale Feature Fusion for Ancient Script Recognition, International Conference on

   Document Analysis and Recognition (ICDAR), pp. 1089- 1094, 2022.

6. P. Gupta, D. Arya, and S. Gavhane, Brahmi Character Recognition Using Convolutional Neural Networks, Jour- nal of Cultural Heritage, vol. 42, pp. 58-67, 2020.

7. M. Jangid and L. Srivastava, Dilated Convolutions for Im- proved Performance in Damaged Text Recognition, IEEE Transactions on Image Processing, vol. 30, no. 5, pp. 1567- 1580, 2021.

8. T. Dhanya and R. Venkatesh, Neural Machine Translation Approaches for Brahmi to Modern Indian Script Conver- sion, Computational Linguistics and Intelligent Text Pro- cessing, pp. 215-227, 2020.

9. S. Malhotra and P. Khare, Context-aware Missing Word Prediction in Ancient Inscriptions, ACM Transactions on Asian and Low-Resource Language Information Process- ing, vol. 19, no. 3, pp. 1-22, 2020.

10. K. Malaviya and S. Chaudhuri, Iconographic Analysis for Contextual Understanding of Ancient Indian Inscriptions, Digital Scholarship in the Humanities, vol. 35, no. 3, pp. 589-603, 2020.

11. R. Patel and A. Mishra, Activation Function Analysis for Ancient Script Recognition, Neural Computing and Appli- cations, vol. 32, pp. 7987-8001, 2020.

12. V. Gupta, S. Reddy, and M. Singh, Residual Connections in CNN for Improved Recognition of Damaged Epigraphic Material, IEEE Access, vol. 9, pp. 45678-45690, 2021.

13. G. Kumar, B. Sharma, and P. Das, Transfer Learning for Low-Resource Ancient Script Recognition, Pattern Recognition, vol. 108, Article 107538, 2020.

14. A. Joshi and S. Patel, Comparative Analysis of OCR Tech- niques for Ancient Indian Scripts, International Journal of Computer Vision and Image Processing, vol. 11, no. 2, pp. 23-41, 2021.

15. R. Tiwari, S. Gupta, and D. Sharma, A Comprehensive Survey of Brahmi Script Recognition: Challenges and Ap- proaches, ACM Computing Surveys, vol. 53, no. 4, pp. 1-34, 2021.

16. M. Khare and R. Sahu, Swish and Mish Activation Func- tions for Damaged Character Recognition, Neural Pro- cessing Letters, vol. 52, pp. 1873-1887, 2020.

17. P. Desai and A. Kumar, Dataset Creation and Augmenta- tion Techniques for Ancient Script Recognition, Data in Brief, vol. 34, Article 106679, 2021.

18. S. Raghunath and T. Jain, Visual-Linguistic Models for Archaeological Inscription Analysis, Journal of Archaeo- logical Science, vol. 128, Article 105345, 2021.

19. Verma et al., Benchmark Database for Brahmi Character Recognition with Degradation Levels, 2021.

20. Mehta and Chowdhury, Complete Processing System for Brahmi Inscription Analysis, 2021.