⚠ Research demo only — not a diagnostic device. This classifier reaches
~87.33% validation accuracy on a held-out set (89.89% test accuracy) and is unsuitable for clinical, medical, or
identity-verification decisions.
Specimen Intake
◇ Insert Fingerprint Image ◇
Drag & drop, or browse files
ACCEPTED: JPEG · PNG · WEBP — max 8MB
Awaiting Specimen
Upload a fingerprint on the left and run classification — the case report will appear here.
Case Report
SPEC-000000
Predicted Blood Group
—
Confidence: —
LowCertaintyHigh
Class Probability Breakdown
Model Details
Architecture
Input 448×448×3
→
conv1–layer2 (frozen)
→
layer3 (fine-tuned)
→
layer4 (fine-tuned)
→
GAP
→
FC 2048→1024→512→128→64→8
Why this shape
- ResNet-50 backbone pretrained on ImageNet; early layers stay frozen since low-level edge/texture filters transfer well to ridge patterns.
- Only
layer3andlayer4are unfrozen — enough capacity to adapt mid/high-level features to fingerprints without overfitting the smaller labeled set. - A tapering FC head (2048→…→8) with BatchNorm + Dropout regularizes the classifier on top of a 2048-d pooled embedding.
- Grad-CAM hooks sit on the last block of
layer4so predictions come with a visual rationale, not just a label.
Reported metrics
87.22%
Val Accuracy
8
Classes
448²
Input Res
Honest limitations
- Fingerprint-to-blood-group correlation is not clinically established; this is a pattern-learning exercise on a research dataset, not a validated biological link.
- Accuracy is reported on a single held-out split — no cross-validation or external test set yet.
- No liveness/spoof detection — the model will happily classify a photo of a photo.
Project Team
Nilotpal Dhar · Koushik Sasmal · Priyadeb Barman · Raj Jaiswal · Monojit Pal
B.Tech Computer Science & Business Systems · Academy of Technology · Final-Year Project