AI paper finds cut-paste synthetic data fails for Indian cheque field localization

Jaganadh Gopinadhan's new paper tackles a practical bottleneck in automated cheque processing: localizing key fields (date, legal amount, IFSC code, account number, signature, and payee name) in Indian bank cheques. The IDRBT Cheque Image Dataset—the only public collection of Indian cheques—lacked field annotations and had unclear redistribution terms. Gopinadhan addresses this by releasing six-field bounding-box annotations for all 112 real cheques (keyed to original filenames to respect IDRBT terms) and generating 295 fully redistributable synthetic images via a cut-paste pipeline that composites real field patches onto content-erased, bank-specific canvas templates. Both resources are available on Hugging Face Hub under permissive licenses.

The baseline test delivers a sobering surprise. A ResNet-50 direct-regression model predicting all six fields in one forward pass was compared against a no-learning baseline that simply predicts each field's mean training box. That trivial baseline already achieves 0.691 mean IoU and 80% accuracy at IoU ≥ 0.5. After accounting for seed variance and training compute, the synthetic-augmented model failed to outperform the real-only model on any aggregate metric—a strong negative result. The paper concludes that appearance-only augmentation is ineffective for fixed-layout documents and argues for layout-varying synthesis. This cautionary finding is fully reported as a reusable resource for the computer vision community.