Aug 2026· African Journal Of Applied Research· Vol 12, pp. 1059-1081· 0 citations· 12 references
TL;DR
A unified multi-attention framework was developed that explicitly optimises structural handwritten character script recognition through an adaptive multi-phase learning-rate schedule, incorporating global and hierarchical local transformer operations (ViT and Swin Transformer), respectively.
Abstract
Purpose: In this study, we explored the effect of multi-scale hybrid deep learning architectures and adaptive optimisation on offline Devanagari handwritten character recognition (HWCR) performance. CNNs are local networks that have been shown to excel at extracting local features but are unable to process the context of complex character configurations.
Design/Methodology/Approach: This paper introduces a unified hybrid multi-attention deep learning framework which systematically combines the complementary merits of CNN layers, the Vision Transformer (ViT), and the Swin Transformer. In conjunction with a sequence-modelling layer based on Bidirectional Long Short-Term Memory (BiLSTM) and a Connectionist Temporal Classification (CTC)- based decoding layer, this visual extraction pipeline eliminates the need for explicit character segmentation. The handwritten Devanagari character dataset from Kaggle was used, but images were resized to 64 x 256 to define the features of input images and define the evaluation benchmarks. To prevent premature changes in the parameters and to pass through plateaus more gracefully, we adopted a dynamic learning rate optimisation scheme based on linear warm-up, cosine annealing, and validation loss.
Research Limitation: Misclassification issues remain for localised areas with highly stylised writing, overlapping strokes, or partially faded handwriting lines.
Findings: The baseline CNN + BiLSTM architecture achieves an accuracy of 93.18%. Incorporating global attention via the Vision Transformer improves accuracy to 95.02%, and adding hierarchical localised window attention through the Swin Transformer increases performance to 96.41%. Introducing the adaptive learning rate strategy yields the final proposed model pipeline, which outperforms all standalone and baseline configurations, achieving a peak classification accuracy of 97.86%, a weighted precision of 0.9829, and a weighted F1-score of 0.9783.
Practical Implication: For the practical implementation and evaluation of the end-to-end framework, we have used the PyTorch deep learning library in a GPU-enabled computing environment.
Social Implication: This research will contribute to improving automated document indexing, archival digitisation, and language accessibility services for widely used scripts in Indian languages.
Originality/Value: A unified multi-attention framework was developed that explicitly optimises structural handwritten character script recognition through an adaptive multi-phase learning-rate schedule, incorporating global and hierarchical local transformer operations (ViT and Swin Transformer), respectively.
This study presents an advanced framework for Telugu handwritten character recognition by integrating Conditional Generative Adversarial Networks (cGANs) with Vision Transformer (ViT) architectures. Critical issues with Telugu scripts, such as intricate character structures, significant inter-writer variability, and a lack of annotated handwritten data, are addressed by the suggested method. While the Vision Transformer utilizes self-attention mechanisms to capture long-range spatial dependencies and global contextual features necessary for accurate recognition, cGAN-based synthetic data augmentation is employed to enhance dataset diversity and mitigate class imbalance. The proposed system outperforms several current CNN-, RNN-, and heuristic-based techniques, achieving character recognition accuracy of 97.89% and word recognition accuracy of 97.34%, as determined through extensive experiments conducted on real and synthetic handwritten datasets. Stable performance under noisy and real-world conditions is further confirmed by robustness analysis. The outcomes confirm the efficacy of integrating transformer-based learning with generative AI, creating a dependable and scalable OCR solution for low-resource Indic scripts, such as Telugu.
Padmavathi Pragada, D. Ch· Engineering Research Express· 0 citations
Offline handwritten Marathi character recognition is still kind of hard research problem because there is so much variability within the same class ,and between classes they can look a bit similar ,also the strokes are complex and different people write in their own style. A lot of CNN and Transformer like methods either do not really capture long range relationships well enough, or they end up being too heavy computationally, you know not so efficient. So in this paper we suggest a Hybrid Vision Mamba and Transformer (HVMT), framework for stronger offline handwritten Marathi character recognition. The HVMT idea combines the hierarchical feature extraction power of Vision Mamba, which uses selective state-space modeling, with the contextual representation learning of a smaller Transformer encoder, and inside that encoder we use Multi-Head Self-Attention. Experiments are done on the public MHCD_GIETV2 dataset, where handwritten Marathi characters are collected from writers in different age groups and with diverse writing styles. Before training the images are turned into grayscale, then normalized, resized, and also augmented, to help the model generalize better. The proposed HVMT is compared with CNN, ResNet-50, EfficientNet-B0, ConvNeXt-Tiny, Vision Transformer (ViT-B/16), Swin Transformer-Tiny, and Vision Mamba, all under the same experimental setup. Experimental results show that the proposed framework achieved accuracy 87.11% , precision 87.09% , recall 87.10% and F1-score 87.09% which is better than the compared architectures. At the same time it only uses 26.9 million parameters, 2.6 GFLOPs, and inference time 1.305 ms per image. In other words, the HVMT framework seems to strike a workable tradeoff between recognition precision and compute efficiency. Because of this it is a good fit for things like intelligent document analysis, handwritten document digitization , archival preservation , and several other Indic script recognition tasks and more.
S. Khandakhani, Sachikanta Dash, Sasmita Padhy et al.· Journal of Intelligent Decis...· 0 citations
Recognition of Indic handwritten text is a difficult issue with the complex formations of characters, variability of graphemes, ambiguity of strokes, and significant differences between writers. This is particularly problematic in scripts (such as Tamil and Kannada) where the form of the handwritten words and the composition structure often are not regular. Although recent CNN-RNN and CNN-Transformer architecture have achieved encouraging results, they either pay much attention to local visual representation or global context representation and do not consider the structural relationship of handwritten patterns at an appropriate level. This work will offer a solution to this drawback by suggesting a Hybrid CNN 10 Capsule 10 Transformer network with Connectionist Temporal Classification (CTC) to perform end-to-end handwritten word recognition. The proposed framework uses CNN layers to extract local visual features, the capsule module to encode structural and compositional relationships, and the Transformer to learn long-range sequence dependencies to be correctly transcribed. The model is tested on Tamil and Kannada handwritten data to check the effectiveness of cross-scripts. The results of the experiment indicate that the given architecture has a test accuracy of 85.46 percent and a Character Error Rate (CER) of 0.0281 on Tamil and 88.70 percent and a Character Error Rate (CER) of 0.0175 on Kannada. These results indicate that the hybrid framework proposed enhances the performance of cross-script handwritten text recognition as compared to baseline architectures.
A S Manjunath, Umesh Dadadahalli Ramu, Madhusudan G et al.· International journal of com...· 0 citations
Offline handwritten text recognition (HTR) remains one of the most challenging problems in document image analysis because of cursive writing, character overlap, unconstrained writing styles, long inter-character dependencies and background degradation. Classical CNN-BLSTM pipelines and single-scale attention models extract features at a fixed resolution and therefore struggle to jointly represent thin strokes of small characters and the global shape of large or connected characters. In this paper we propose a Multi-Scale Transformer-Based Lexicon-Guided HTR framework built around an Adaptive Feature Fusion (AFF) mechanism. A multi-scale convolutional backbone extracts shallow, middle and deep feature maps using parallel 3×3, 5×5 and 7×7 receptive fields. The AFF module learns content-dependent soft weights and channel attention to fuse these heterogeneous features into a single scale-balanced representation, replacing the fixed feature map used by previous lexicon-and-attention systems. A Transformer encoder with multi-head self-attention then models long-range context far more effectively than bidirectional recurrent layers. The network is trained end-to-end with the Connectionist Temporal Classification (CTC) objective and decoded with a lexicon-guided beam search that combines an n-gram prior with edit-distance dictionary matching. On the IAM line-level benchmark the proposed model attains a character error rate (CER) of 3.08% and a word error rate (WER) of 7.24%, improving over our lexicon-and-attention baseline (4.15% CER / 9.72% WER) by 25.8% and 25.5% relative, respectively. On RIMES we obtain 2.71% CER / 7.92% WER, and on the George Washington collection 5.82% CER / 13.05% WER, again outperforming the baseline.
Lalita Kumari· International journal on eme...· 0 citations
This paper introduces NepScript Genesis, a Neural Architecture Search (NAS) framework for automated Generative Adversarial Network (GAN) discovery, applied to conditional Devanagari handwritten digit synthesis. We compare five NAS strategies against a carefully constructed Deep Convolutional GAN (DCGAN) baseline (FID=332.28). Architecture selection utilizes a two-stage pipeline guided by a novel domain-aware evaluation metric (Enhanced Score). Results demonstrate that Adaptive Exploration achieves the optimal quality-efficiency trade-off, attaining an FID of 79.12 -- a 76.19% improvement over the baseline -- and the highest mode coverage among the NAS strategies (Recall=0.531) in under one GPU-hour. Furthermore, we demonstrate that incorporating script-specific structural heuristics into the search phase prevents early-stage mode collapse. In a downstream low-resource evaluation, augmenting 250 real training samples per class with GAN-generated digits from the best NAS model improves CNN classification accuracy from 91.0% to 96.5% (+5.5 percentage points), demonstrating that NAS-optimized synthesis produces digits of sufficient quality to benefit practical recognition pipelines when real data is scarce.
Mausam Gurung, Prabin Neupane, S. Acharya· 0 citations
Vision Transformers (ViT) capture global image context through self-attention but are data-hungry, typically underperforming Convolutional Neural Networks (CNNs) on the small datasets common in fine-grained tasks such as batik motif recognition. This study investigates whether a ViT, trained via knowledge distillation using the Data-efficient Image Transformer (DeiT), can overcome this limitation and compete with CNNs on a small Semarang Batik dataset. A distilled DeiT-Tiny student learns from a ResNet-50 CNN teacher and is benchmarked against two CNN references: ResNet-50 itself (a substantially larger model) and EfficientNet-B0 (a parameter-matched counterpart). In establishing this comparison, we first uncover a critical dataset integrity issue: the publicly available Semarang Batik Dataset (3,020 images) originates from only 18 unique source photographs, each augmented approximately 167 times prior to publication. This near-duplication causes severe data leakage under conventional random splitting, inflating the test accuracy of all models to a misleading 100% and rendering such evaluation meaningless. We therefore introduce a source-aware splitting strategy that enforces group-level separation between training, test partitions, and evaluate all models across three random seeds for statistical reliability. Under this corrected protocol, the distilled DeiT-Tiny attains the highest mean accuracy (95.18 ± 0.30%) and the lowest variance among the three models, matching both the larger ResNet-50 (94.87%) and the parameter-matched EfficientNet-B0 (94.68%) while using only 5.5M parameters. These results confirm knowledge distillation enables a compact Vision Transformer to compete CNNs on a limited fine-grained dataset, and underscore that verifying sample independence is a prerequisite for trustworthy evaluation on pre-augmented public datasets.
Rafi Alifa Bagja, Bedy Purnama· International Conference on...· 0 citations