BanglaHand: A Dataset of Handwritten Paragraphs for Writer-Independent Handedness Recognition
Description
BanglaHand is a benchmark dataset of 42 writers (21 naturally left-handed and 21 right-handed) designed for offline handwriting analysis and handedness recognition in Bangla script. It contains 336 high-resolution smartphone photographs of complete handwritten paragraphs captured under rigorously controlled conditions. Key Highlights: • Balanced Demographic Benchmark: 21 confirmed left-handed and 21 right-handed writers, providing a 1:1 balanced distribution in a population where left-handed writing is exceptionally rare (approx. 1–2%). • Dual-Condition Text Design: Each writer contributed 8 paragraphs: - 6 shared prompts (S01–S06, 49–54 words each) identical across all writers for content-controlled benchmarking. - 2 writer-specific paragraphs on self-chosen topics to test generalization to unseen text. • Controlled Materials & Capture: All participants wrote using pens and paper from a single common supply in mixed writing sessions. All pages were photographed in one session with a Samsung Galaxy A17 smartphone (4080 × 2296 pixels) in mixed order. • Geometric Audit & Quality Control: A SIFT + RANSAC homography screening across all writer pairs confirmed writer uniqueness and prevented page duplicates. Images were rotated upright (lossless 4:4:4 JPEG, quality 98) without cropping, resizing, or photometric filtering. Archive Structure: • Left-Handed/ : 168 paragraph images across 21 writers (P1–P21) • Right-Handed/ : 168 paragraph images across 21 writers (P1–P21) • metadata/images.csv : Image-level records (capture timestamps, SHA-256 hashes, camera EXIF, and quality flags) • metadata/prompts.csv : Ground-truth Bangla text, English translations, and word counts for S01–S06 • metadata/writers.csv & excluded_writers.csv : Writer profiles and exclusion provenance • SHA256SUMS.txt & README.md : Byte-level integrity checksums and full data documentation Research Applications: • Writer-Independent Handedness Identification (Soft Biometrics) • Forensic Document Examination & Questioned Handwriting Profiling • Candidate Search-Space Pruning in Automated Writer Identification • Clinical Neuromotor Lateralization & Motor Control Studies • Indic Script Offline Handwritten Text Recognition (HTR) • Audits of Acquisition Bias & Shortcut Learning in Deep Learning Companion Publications: 1. Data Descriptor: "BanglaHand: A Dataset of Handwritten Paragraphs for Writer-Independent Handedness Recognition", Data in Brief (submitted). 2. Methodology: "Handedness identification from Bangla handwriting: A writer-independent evaluation across text conditions", Pattern Recognition Letters (submitted).
Files
Steps to reproduce
1. Download and extract 'bangla_handedness_paragraphs_v1.zip'. 2. Open 'metadata/images.csv' to view the full manifest of 336 paragraph images, writer IDs (L_p01–L_p21, R_p01–R_p21), handedness labels (0 = right, 1 = left), capture timestamps, and condition tags. 3. For content-controlled writer-independent benchmarking, filter by condition == 'shared' (paragraphs 1–6, prompts S01–S06). For open-vocabulary generalization testing, use paragraphs 7–8. 4. Thirteen images that cut the last line at the bottom frame edge are flagged in the 'text_truncated_at_edge' column for optional filtering. 5. Verify archive integrity using the provided SHA-256 hashes in 'SHA256SUMS.txt'.
Institutions
- Shahjalal University of Science and TechnologySylhet Division, Sylhet