Devanagari Handwritten Character Dataset (DHCD)

• DHCD is an open benchmark dataset for offline handwritten recognition in Devanagari, offering versatile configurations for numeral and character studies.
• It employs detailed acquisition and preprocessing protocols, including controlled segmentation and quality-check procedures across multiple corpus versions.
• The dataset has enabled advancements in deep CNN and hybrid quantum-classical models, with reported accuracies up to 99.80% in digit recognition.

The Devanagari Handwritten Character Dataset (DHCD), also spelled “Devnagari Handwritten Character Dataset” in earlier work, is an open benchmark line for offline handwritten recognition in the Devanagari script. In the cited literature, the acronym refers not to a single invariant corpus specification but to related dataset releases, curated subsets, and task-specific interpretations used for isolated numeral recognition, isolated character recognition, and digit-only benchmarking. Its documented roles include benchmark database generation, comparative evaluation of feature extractors and classifiers, and more recent evaluation of CNN-based and hybrid quantum-classical recognition systems (Dongre et al., 2013, Mehta et al., 14 Jul 2025, Malla, 8 Jun 2025).

1. Origin and benchmark role

The earliest cited benchmark formulation is the database-generation study by Dongre and Mankar, which presents DHCD as an offline benchmark database for Devnagari handwritten numerals and characters. That work reports 5,137 isolated samples for the numeral database and 20,305 isolated samples for the character database, collected from 750 writers of all ages, sex, education, and profession. The stated motivation is the lack of an official benchmark for Devnagari script and the need for free access for researchers working on handwritten character recognition (Dongre et al., 2013).

In that original description, the dataset is positioned as an open-access “MNIST-style” resource for Devnagari, intended to enable apples-to-apples comparison across feature-extraction methods such as Zernike moments, chain profiles, SIFT, and HOG, and across classifiers such as k-NN, SVM, and neural nets. The original paper does not report recognition rates; it is explicitly a database-generation paper rather than a recognition paper. Subsequent works are described as using DHCD to report accuracies in the 85–95% range depending on feature extractor and classifier (Dongre et al., 2013).

Later papers repurpose the name DHCD in narrower and more model-centric ways. A 2025 CNN paper describes DHCD as an open dataset for Devanagari handwritten text recognition and uses it for handwritten character classification, while a 2025 hybrid quantum-classical paper uses a digit subset from a corpus it also calls DHCD for Devanagari digit recognition (Mehta et al., 14 Jul 2025, Malla, 8 Jun 2025).

2. Reported corpus configurations

The literature summarized here reports multiple incompatible but related DHCD configurations.

Source	Reported scope	Reported counts
Dongre and Mankar (Dongre et al., 2013)	Numerals and characters	5,137 numerals; 20,305 characters
CNN study (Mehta et al., 14 Jul 2025)	Devanagari characters	36 classes in the abstract; 47 “main” letters in the dataset overview; 1,700 images per class
Quantum study (Malla, 8 Jun 2025)	Digits plus consonants	92,000 grayscale samples overall; digit subset of 20,000 images

The 2013 database description specifies 10 numeral classes for Devnagari digits $०$ through $९$, and 46 core Devnagari character classes, including 13 vowels/modifiers and 33 consonants/compound forms, plus a special folder for the “@” symbol, for a total of 50 character folders (Dongre et al., 2013).

The 2025 CNN study presents a different specification. Its abstract states that DHCD is “an open dataset with 36 classes of Devanagari characters,” with each class having 1,700 images for training and testing purposes. Its detailed dataset overview, however, says that DHCD is built upon 47 “main” letters of the Devanagari script: 14 vowels and 33 consonants. The same overview states that these 47 character-classes form the label set for all handwritten samples, with approximately 61,000 scanned and clipped character images, while also noting that $1{,}700 \times 47 \simeq 79{,}900$ and that the paper rounds to 61,000 overall to reflect usable images after clipping and quality control (Mehta et al., 14 Jul 2025).

The 2025 quantum study defines DHCD differently again, reporting 92,000 grayscale samples overall, composed of 10 digit classes and 36 consonant classes. Its experiments use only the digit subset: 20,000 images across ten classes, with exactly 2,000 samples per digit and no class imbalance (Malla, 8 Jun 2025).

A common misconception is that DHCD has a single canonical class inventory and sample count. The cited sources do not support that assumption. This suggests that later papers may rely on curated subsets, task-specific extractions, or differently defined releases rather than a single immutable corpus specification (Dongre et al., 2013, Mehta et al., 14 Jul 2025, Malla, 8 Jun 2025).

3. Acquisition, annotation, and file representation

The original database-generation paper gives the most detailed acquisition protocol. It reports 750 distinct writers drawn from schoolchildren through seniors, with both male and female participants and a heterogeneous population including students, office staff, factory workers, housewives, and senior citizens. Writers used ball-point, gel, or ink pens, with no restriction on instrument. Samples were written on standardized A4 sheets printed with empty rectangular boxes plus a vertical guide line in each row. The sheets were scanned using a Canon CanoScan LiDE 100 flatbed scanner at 300 dpi, and the raw scans were saved as JPEG before automated symbol extraction (Dongre et al., 2013).

In the 2013 formulation, final symbol images are cropped to their tight bounding box and stored in TIFF format at binary black-and-white pixel depth. The paper gives the storage formula

$\text{Size (bytes)} = \frac{W \times H}{8},$

with the example that a $32 \times 32$ binary image requires 128 bytes. The dataset is organized into class-named directories, including /numerals/0 through /numerals/9 and /chars/अ/ through /chars/ज्ञ/, plus /chars/@/. Each image receives a unique sequential identifier such as 0001.tif, and a separate metadata.txt maps image index to writer ID and class label (Dongre et al., 2013).

The later CNN paper describes a distinct character-collection process. It states that hundreds of handwritten pages were collected from multiple authors to capture a wide variety of stroke widths, slants, and personal styles. Each character was manually cropped from scanned pages, with care taken to include only clear, single-character crops, and every image was labeled according to its corresponding Devanagari letter, whether vowel or consonant. The paper also highlights high intra-class variability and notes visually similar pairs such as $प$ vs. $य$ and $घ$ vs. $ध$, which increase the risk of mislabeling and require precise annotation (Mehta et al., 14 Jul 2025).

The quantum study is more compact in its data-description layer. For the digit subset, it specifies $28 \times 28$ single-channel grayscale images with intensity in $९$0, and states that no additional metadata was provided or used (Malla, 8 Jun 2025).

4. Preprocessing, curation, and split protocols

The 2013 DHCD preprocessing pipeline is explicitly automated in MATLAB routines. The documented stages are: grayscale-to-binary conversion via global thresholding; morphological cleaning for noise removal of isolated pixels; border and guide-line removal by detecting the largest connected component in each scan; row segmentation through horizontal projection histograms, where zero runs indicate row gaps; column segmentation by bounding boxes of contiguous pixel groups while ignoring the leftmost guide line; and cropping and saving each symbol as a separate TIFF file. Ambiguous or poor-quality segments, including stray strokes and over- or under-segmented shapes, were discarded by visual checking (Dongre et al., 2013).

The 2013 paper does not define a fixed train/test split. It states that the authors intend to categorize training and test sets randomly in the near future. The summary notes that typical follow-up practice is to partition each class into 70% training, 15% validation, and 15% test, either randomly or stratified by writer, but this is not an original DHCD protocol (Dongre et al., 2013).

The 2025 CNN study describes a more image-normalized character pipeline. Original clipped characters occupy a $९$1-pixel box, and each sample is stored as a $९$2-pixel image by adding 2-pixel zero-padding on all four sides. The preprocessing steps are grayscale conversion from RGB to a single intensity channel, intensity inversion from black-on-white to white-on-black, and background suppression in which all non-foreground pixels are set to zero intensity. The paper states that no explicit normalization formula such as $९$3 is provided (Mehta et al., 14 Jul 2025).

That same study reports a nominal 80%/20% train/test split. With 1,700 images per class, the derived counts are 1,360 training images and 340 testing images per class. At the aggregate level, it gives $९$4 training images and $९$5 testing images, while also stating that these are reported as 48,960 and 12,240 after excluding low-quality scans. It further notes that no on-the-fly or offline augmentation such as rotation, scaling, or shearing is reported; all experiments use only raw clipped and preprocessed images (Mehta et al., 14 Jul 2025).

The 2025 quantum study adopts a separate, digit-only protocol. It applies the normalization

$९$6

so that $९$7. Training-only augmentations are random rotation with $९$8, affine translation with $९$9, horizontal flip with $1{,}700 \times 47 \simeq 79{,}900$0, and elastic distortion with $1{,}700 \times 47 \simeq 79{,}900$1, $1{,}700 \times 47 \simeq 79{,}900$2, and $1{,}700 \times 47 \simeq 79{,}900$3. For $1{,}700 \times 47 \simeq 79{,}900$4 digit samples, the split is 17,000 images for training plus validation and 3,000 for test, with the former divided into 13,600 training and 3,400 validation images; the split is stratified so that each class has 1,360 train and 340 validation samples. Validation and test images undergo normalization only (Malla, 8 Jun 2025).

5. Use in recognition studies

In contemporary recognition papers, DHCD functions not merely as an archival corpus but as an evaluation substrate for increasingly specialized architectures.

The CNN-based character-recognition study frames its task as automated extraction of handwritten Hindi characters from an image of Devanagari script. It presents a recognition technique using two deep convolutional neural network layers and states that the configuration is intended to enhance the recognition rate for Devanagari handwritten text recognition (DHTR). On the reported DHCD setup, it achieves 96.36% accuracy in testing and 99.55% in training time (Mehta et al., 14 Jul 2025).

The hybrid quantum-classical study narrows the problem to handwritten Devanagari digits. It uses a CNN for spatial feature extraction and a 10-qubit variational quantum circuit for classification. On the DHCD digit subset, it reports 99.80% overall test accuracy, corresponding to 2,994 correct predictions out of 3,000; a test loss of 0.2893; and an average per-class F1-score of 0.9980. The study also reports a classical CNN baseline with 99.03% accuracy and 0.9912 average F1, and notes an accuracy improvement of approximately 0.77 percentage points for the hybrid model (Malla, 8 Jun 2025).

The same quantum study gives a precise task formulation around balanced digit data. It defines test accuracy as

$1{,}700 \times 47 \simeq 79{,}900$5

and reports that most of the few misclassifications arise among visually similar digits such as “3” vs “6” and “5” vs “2.” It attributes improved generalization to augmentation, including rotation and elastic distortion, and to the variational quantum circuit’s entanglement structure (Malla, 8 Jun 2025).

These studies show that DHCD is used across materially different task regimes: isolated characters without reported augmentation, and digit-only experiments with aggressive augmentation and normalized grayscale inputs. This suggests that direct comparison of headline accuracies across DHCD papers requires attention to the exact subset, class inventory, curation policy, and preprocessing pipeline employed (Mehta et al., 14 Jul 2025, Malla, 8 Jun 2025).

6. Access, licensing, and recurrent ambiguities

The original 2013 database description provides the clearest access and licensing terms. It states that the dataset is publicly available at http://code.google.com/p/devnagari-database/ and released under the Creative Commons Attribution License, permitting research use, redistribution, and modification provided that the original citation is maintained (Dongre et al., 2013).

The 2025 CNN paper also describes DHCD as an “open dataset,” but explicitly states that no formal license such as Creative Commons is specified in the paper. For access, it points readers to reference [12], identified as Acharya, Pant, and Gyawali (2015), “Deep learning-based large scale handwritten Devanagari character recognition,” SKIMA 2015, for download links and data-usage terms (Mehta et al., 14 Jul 2025).

The quantum study does not focus on licensing. Its principal limitation statements concern experimental scope: only the digit subset of 20,000 samples is used; the full DHCD, including 36 consonant classes, remains to be explored; the quantum layers are simulated in PennyLane; real-hardware noise may degrade performance; and scalability beyond 10 qubits is limited by current NISQ hardware (Malla, 8 Jun 2025).

Several ambiguities recur across the literature. One concerns script spelling, with “Devnagari” used in the 2013 source and “Devanagari” in the 2025 papers. Another concerns class cardinality, with reported inventories of 36 classes, 47 “main” letters, 10 digits plus 36 consonant classes, and 46 core characters plus an “@” folder. A third concerns total sample count, ranging from 25,442 isolated numeral-plus-character samples in the original release to approximately 61,000 curated character images in one later description and 92,000 grayscale samples in another. A plausible implication is that “DHCD” has become a literature-level umbrella label for related Devanagari handwritten benchmarks and subsets rather than a single frozen artifact (Dongre et al., 2013, Mehta et al., 14 Jul 2025, Malla, 8 Jun 2025).

Despite these inconsistencies, the dataset family remains central because it fills a long-standing gap for open-access offline handwritten recognition in Devanagari, supports both numeral and character classification, and exposes the core technical challenges of the script: high intra-class variability, visually similar classes, and sensitivity to segmentation and annotation quality (Dongre et al., 2013, Mehta et al., 14 Jul 2025, Malla, 8 Jun 2025).