CEDAR Handwriting/Signature Datasets
- CEDAR Handwriting Dataset is a curated collection of 'AND' word fragments from over 1,500 writers, enabling precise writer verification.
- CEDAR Signature Dataset comprises 24 genuine and 24 forged signatures per signer to support robust offline signature verification.
- Both datasets use strict acquisition, preprocessing, and writer-disjoint splits to ensure fair benchmarking of classical and deep learning methods.
The Center of Excellence for Document Analysis and Recognition (CEDAR) at the University at Buffalo is internationally recognized for producing benchmark datasets that have catalyzed progress in offline handwriting analysis and signature verification. The CEDAR datasets encompass distinct collections for handwriting and signature analysis, each defined by rigorous acquisition protocols, systematic annotation, and widespread adoption in the evaluation of both handcrafted and deep learning-based approaches to writer and signer identification.
1. CEDAR Handwriting Dataset: The "AND" Subset
The CEDAR handwriting dataset is derived from the CEDAR Letter corpus, which contains manuscript letters written in English by distinct writers. Each participant contributed three full-page letters written on standard white letter-size paper with their preferred black ink pen, under typical indoor office lighting. Data acquisition was performed via the CEDAR-FOX imaging system at 600 dpi, maintaining consistency with signature dataset standards (Chauhan et al., 2024). From each letter, up to five occurrences of the word “AND” were identified and extracted using a transcript-mapping algorithm embedded in CEDAR-FOX. This segmentation process yielded a total of 15,518 tightly-cropped word image fragments across all writers, averaging approximately 9.9 “AND” snippets per writer.
No signature images from the broader CEDAR project were utilized in this subset. The focus is exclusively on repeated handwriting fragments, supporting studies in writer verification through isolated lexical content (Chauhan et al., 2024).
2. CEDAR Signature Dataset
The CEDAR signature dataset targets offline signature verification. It consists of contributions from 55 signers, with each providing 24 genuine signatures and 24 corresponding skilled forgeries, resulting in $1,320$ genuine and $1,320$ forged signature images (Chokshi et al., 2023). Images were scanned and resized to pixels, then converted to grayscale. No additional morphological normalization or binarization was reported prior to feature extraction. The dataset partitions 40 writers (1,920 images) for training and 15 writers (720 images) for testing, aligning with common cross-writer evaluation protocols. No explicit validation set or data augmentation was documented (Chokshi et al., 2023).
A summary comparison is given below.
| Dataset | Writers | Samples per Writer | Use Cases |
|---|---|---|---|
| CEDAR "AND" (Handwriting) | 1,567 | ≈10 (“AND” fragments) | Writer verification, SSL |
| CEDAR Signature | 55 | 24 genuines, 24 forgeries | Signature verification, Siamese |
3. Preprocessing, Partitioning, and Sampling Protocols
For the "AND" collection, images underwent padding (preserving aspect ratio), resizing to pixels, and foreground inversion (white strokes on black) (Chauhan et al., 2024). Feature extraction for baseline comparisons involved binarized images passed through two primary pipelines: GSC (512-dimensional global shape context) and OpenCV’s HOGDescriptor (1,764-dimensional histogram of oriented gradients).
Contrastive self-supervised learning (CSSL) and related deep learning approaches employed a robust augmentation pipeline, including random crop and resize, horizontal and vertical flips, small angle rotations, perspective warping, Gaussian blur, jittering, and inversion. These augmentations aim to preserve writer-specific micro-features while inducing invariance to nuisance variables.
For downstream writer verification, binary (anchor, query) pairs are constructed with balanced genuine (same-writer) and impostor (different-writer) sampling: . Writer-independent data splits are mandatory. In (Chauhan et al., 2024), writers to () define the training cohort for both pre-training and supervised fine-tuning, while the remaining $367$ writers are reserved for test evaluation—a strict $1,320$0 policy.
Fine-tuning is performed under two annotation regimes: (i) using only $1,320$1 of the train writers (120 writers, 13,232 balanced pairs), and (ii) all $1,320$2 train writers (129,602 pairs). Supervised fine-tuning utilizes small held-out validation subsets for early stopping based on the $1,320$3 score (patience=5, $1,320$4). No cross-validation is performed; all results reflect this canonical train/test split (Chauhan et al., 2024).
In the signature dataset, training and testing are also partitioned at the writer level, with 40/15 writers defining train/test splits. Images are resized and converted to grayscale, with no report of binarization or further normalization before feature extraction or classification. Sampling for verification tasks strictly respects the signer-disjoint splits.
4. Evaluation Metrics and Quantitative Criteria
For handwriting verification with the "AND" dataset, self-supervised representation learning techniques are evaluated initially using a cosine-separation metric. Given embeddings $1,320$5 and $1,320$6, cosine similarity is given by: $1,320$7 Mean intra-writer and inter-writer similarities are computed; higher $1,320$8 separation indicates superior feature discriminability and aligns with increased downstream verification accuracy (Chauhan et al., 2024).
In signature verification tasks, a Siamese network computes 128-dimensional embeddings for input images; the Euclidean distance $1,320$9 is the principal similarity measure. Verification is thresholded at the Equal Error Rate (EER) operating point, where false acceptance and false rejection rates are balanced. Empirical performance on CEDAR signatures includes:
- ROC AUC: 99.9%
- PR AUPR: 99.9%
- EER: $1,320$0 (surpassing prior art) (Chokshi et al., 2023)
Performance on the "AND" dataset for self-supervised approaches includes a ResNet-based VAE achieving $1,320$1 accuracy and ResNet-18 fine-tuned with VICReg reaching $1,320$2 accuracy. Deployment of pre-trained VAE and VICReg yields relative improvements of $1,320$3 and $1,320$4 over a supervised ResNet-18 baseline with $1,320$5 label coverage (Chauhan et al., 2024).
5. State-of-the-Art Architectures and Feature Extractors
Handcrafted baselines for handwriting data include GSC and HOG-based pipelines, while supervised and self-supervised deep learning leverage convolutional backbones (ResNet variants), variational autoencoders, and contrastive regularization frameworks such as VICReg. For signature verification, SigScatNet introduces a Siamese network with a front-end scattering wavelet transform. The scattering module, with maximum scale $1,320$6, computes translation-invariant and deformation-stable coefficients using Morlet wavelet banks (6–8 orientations). Each Siamese branch consists of four convolutional blocks and a final dense layer mapping $1,320$7 grayscale input to $1,320$8 (Chokshi et al., 2023).
SigScatNet demonstrates that incorporating scattering transforms allows the use of shallow, low-parameter convolutional architectures (approximately $1,320$9 parameters per branch, 0 MB in 32-bit weights). The removal of deep convolutional layers and dimensionality reduction improves computational efficiency—making real-time inference feasible on commodity CPUs with negligible accuracy tradeoff (Chokshi et al., 2023).
Loss functions for training include the triplet loss: 1 where 2 are anchor, positive, and negative samples, and 3 is the margin parameter.
6. Applications, Recommendations, and Future Development
The CEDAR handwriting and signature datasets are foundational in benchmarking handwriting and signature verification algorithms for real-world deployment. Applications include personal identification, forensic document examination, and access control. The structure and strict acquisition and sampling protocols support both classical (handcrafted feature engineering) and contemporary (self-supervised, deep metric learning) research agendas.
Practitioners are encouraged to exploit wavelet scattering transforms to generate robust, multi-scale descriptors while constraining network capacity to match onboard and edge compute limitations (Chokshi et al., 2023). The CEDAR "AND" subset is especially suitable for studying writer verification with minimal lexical content, providing a stressful case for both supervised and self-supervised representation learning (Chauhan et al., 2024).
Integration of online augmentation (rotations, elastic deformations) could further enhance generalization to variant acquisition conditions. For both handwriting and signature tasks, strict writer/signer-disjoint train/test splits are recommended to guard against overfitting, in line with the best practices exemplified in recent work.
7. Limitations and Common Misconceptions
It is often assumed that the CEDAR datasets constitute a monolithic resource; in practice, they are distinct and tailored for specific tasks—offline handwriting or signature verification. For instance, the "SSL-HV" study exclusively employs handwriting snippets and deliberately excludes all signature data, whereas signature-forgery works such as SigScatNet use only the genuine and forged signature images (Chauhan et al., 2024, Chokshi et al., 2023). The consistent use of carefully-balanced, writer-independent splits is crucial for a fair assessment; singlefold evaluation (rather than k-fold cross-validation) is standard in the outlined studies.
A plausible implication is that results derived from one CEDAR subset cannot be readily compared or transferred to the other without regard to composition, acquisition protocols, and task specificity. For both datasets, absence of online data augmentation or explicit validation folds should be considered when interpreting reported metrics.
References: