BingCheck Dataset for Signature Segmentation

• BingCheck Dataset is a curated corpus of high-resolution bank check images featuring dual-layer (pixel and patch) signature annotations for precise segmentation.
• It integrates images from public datasets, internet searches, and original scans to capture diverse real-world check conditions.
• The dataset supports comprehensive evaluation using pixel-level metrics, offering a reproducible testbed for document analysis and signature extraction research.

The BingCheck Dataset (BCSD) is a curated corpus of high-resolution scanned and photographic images of bank checks featuring handwritten signatures accompanied by both pixel-level and patch-level ground truth segmentation masks. Designed for the evaluation and development of signature extraction models and complex document segmentation algorithms, BCSD addresses the challenge of reliably delineating foreground handwritten content—specifically signatures—from intricate and variable backgrounds. This resource provides a unique real-world testbed for machine learning methods requiring fine-grained annotation and is openly accessible, with comprehensive metadata, to facilitate reproducibility and benchmarking in document analysis research (Khan, 2021).

1. Composition, Sources, and Dataset Statistics

BCSD consists of 158 JPEG images of bank checks, standardized such that each image’s longer dimension is precisely 2240 pixels at 300 dpi. The data collection integrates multiple sources:

• IDRBT Cheque Image Dataset: 110 images (≈70%) originate from this publicly available corpus.
• Internet Searches: 37 images (≈23%) were collected via Google Image Search, including counterfeit and template checks.
• Author-Generated Scans: 9 images (≈7%) stem from scans or camera captures of real or dummy checks.

Each check contains one or occasionally more signatures, with disjoint bounding-boxes for distinct instances and merged boxes for overlapping cases. The spatial occupancy of signatures is consistently minor—typically under 5% of total image pixels—enforcing a pronounced class imbalance between foreground (signature) and background. Such statistics are material for both segmentation metric design and the interpretation of model performance, particularly in low-prevalence foreground tasks.

2. Annotation Protocol: Manual and Automated Mask Generation

BCSD provides two distinct layers of signature annotation:

• Pixel-Level Masks: Manual annotation utilized GIMP; the operator superimposed the source RGB check onto a black canvas of identical dimensions, meticulously tracing signature pixels in white (255), before saving single-channel, black-and-white JPEG masks.
• Patch-Level Masks: An automated Python+OpenCV script processes the manual mask by identifying all white pixels and computing the minimal bounding rectangle for each signature. Where bounding rectangles of signatures overlap, a single merged patch is produced; otherwise, boxes remain disjoint. The rectangle is rasterized onto a mask using OpenCV routines.

Quality control involved overlay-based visual inspection and manual correction of discrepancies such as stray or missing strokes. Verification of multi-signature cases ensured accurate bounding for each disjoint region. This two-tier annotation supports both fine-grained and region-based segmentation research.

3. Organization, Partitioning, and Repository Structure

The dataset is partitioned into a training set of 129 images and a test set of 27 images; no validation split is prespecified, but users may carve out such a set from the training partition as required. Directory hierarchy is as follows:

root/ │ ├─ train/ │ ├─ X/ # Original RGB images (JPEG) │ └─ y/ │ ├─ pixel/ # Pixel masks │ └─ patch/ # Patch masks └─ test/ (as above)

Within each split, corresponding masks maintain base filename parity with the original image, simplifying automated pairing during experiments. The repository includes a README specifying source attribution, licensing, file counts, and curation methodology (Kaggle DOI: 10.34740/KAGGLE/DS/1041694).

4. Mask Formats and Interoperability

Segmentation masks in both pixel and patch format are stored as single-channel (grayscale) JPEG images with binary encoding: black ($=0$) corresponds to the background, and white ($=255$) denotes signature pixels or patch regions. Conventional computer vision libraries are recommended for reading these masks:

• OpenCV: cv2.imread(path, cv2.IMREAD_GRAYSCALE)
• PIL/Pillow: Image.open(path).convert('L')
• scikit-image: skimage.io.imread(path, as_gray=True)

The consistent mask format supports seamless integration into deep learning pipelines and cross-library interoperability, facilitating comparison across experimental setups.

5. Evaluation Metrics and Benchmarking Protocol

Evaluation follows standard binary segmentation metrics:

• Intersection over Union (IoU): $IoU = |P \cap G| / |P \cup G|$, with $P$ as the predicted mask and $G$ as the ground-truth mask.
• Pixel Accuracy: $(TP + TN)/(TP + FP + TN + FN)$
• Precision: $TP/(TP + FP)$
• Recall: $TP/(TP + FN)$

No quantitative segmentation baselines are reported by the dataset authors; the primary contribution centers on dataset release and supporting infrastructure. This suggests users must establish their own empirical baselines for fair assessment on BCSD.

6. Intended Applications and Modeling Recommendations

Recommended baseline architectures include U-Net, Fully Convolutional Networks (FCN), DeepLab, and SegNet, or any modern encoder–decoder approach suitable for binary document segmentation. Given large image size and fine-grained signature traces, patch-based training or multi-scale inputs are advised to mitigate downsampling artifacts and preserve signature intricacies. Data augmentation—specifically geometric and photometric transformations—is encouraged to promote robustness to acquisition conditions. Training a U-Net using image crops of up to $2240 \times N$ is feasible within a few hours on a single NVIDIA-class GPU with 12 GB VRAM. Use cases include:

• Preprocessing for signature verification pipelines.
• Document layout analysis in the presence of complex, non-uniform backgrounds.
• Benchmarking in class-imbalanced, foreground–background segmentation.

A plausible implication is that BCSD is uniquely positioned for research into instance segmentation under real-world high-imbalance conditions and non-uniform document artifacts.

7. Accessibility, Limitations, and Research Impact

BCSD is openly available through Kaggle, with detailed metadata and licensing. Its blend of real, synthetic, and internet-derived checks, standardized imaging conditions, and rigorous annotation protocol render it a reproducible and adaptable resource for the document analysis and signature extraction communities. The small foreground-to-background ratio, source heterogeneity, and dual-layer ground truth annotation make BCSD especially pertinent for research striving toward generalized, high-fidelity document segmentation under adverse imaging and layout conditions (Khan, 2021).