All Case Studies
Computer Vision
Signature Recognition using CNN
Comparing CNN feature learning against classical HOG / SIFT for signatures.
2025
OVERVIEW
The Project
A signature-recognition study that processes signature images, extracts features, and classifies signatures by individual ID. It benchmarks CNN-based feature extraction against traditional techniques (HOG and SIFT), with a full segmentation and train/test pipeline.
Objectives
Tools & Technologies
PythonCNNHOGSIFTOpenCV
FIGURES
Charts & Graphs
METHODOLOGY
The Approach
1
Collect a labelled signature dataset (16 images × 12 rows × 4 signatures).
2
Preprocess: normalize, reduce noise, segment signatures from background.
3
Train a CNN (conv + pooling + dense) and evaluate with accuracy, precision, recall and F-measure.
4
Generate confusion matrices for HOG and CNN.
OUTCOME
Results & Learnings
Key Learnings
- Signature-style variability demands robust preprocessing.
- Classical descriptors remain strong baselines on tiny datasets.
