Signature validation using KNN

Vietnamese version here

Idea

• Convert each image into histogram, with adjustable bins.
  • From 1 photo, turn 45 degrees and 90 degrees respectively to have 2 other images. Gather up into a bigger image.
  • Cut the picture vertically, the number of pieces cut by the number of bins of the histogram.
  • For each cut of the image, calculate the total value of the pixels containing the image signature. This sum will be the value of each bins.
• Convert histogram to vector. This vector will be the input of the algorithm.
• Use the built-in k-nearest neighbors (KNN) algorithm of the Sklearn library to trainning and predict

The process

• Set parameters when testing.
  • Standardized image file size: default is (600,400)
  • Number of bins of histogram: default is 100 bins.
  • K: default is 3.
• Get trainning and test data, with each image converted to histogram.
  • Pretreatment
    • Retrieve image data
    • Eliminate only noise and standardize image files.
    • Convert images to histograms for trainning and test data.
• Perform trainning and prediction:
  • Sequentially change the input parameters to statistic accuracy.
• Calculate accuracy measurement.

Trainning and testing

Perform trainng and predict in different parameters.
Based on the statistics below, we see the highest accuracy > 96% when using 20 bins histogram for most image sizes and number of K-neighbor.

Statistic

Image size: (600, 200)
K=5
bins=[600, 500, 400, 300, 200, 100, 90, 80, 70, 60, 50, 40, 30, 20, 10]

Bin	Accuracy
600	63.9175257732
500	67.0103092784
400	76.2886597938
300	70.1030927835
200	69.0721649485
100	77.3195876289
90	78.3505154639
80	78.3505154639
70	80.412371134
60	84.5360824742
50	87.6288659794
40	90.7216494845
30	95.8762886598
20	96.9072164948
10	95.8762886598

Image size: (600, 200)
K=3
bins=[600, 500, 400, 300, 200, 100, 90, 80, 70, 60, 50, 40, 30, 20, 10]

Bin	Accuracy
600	70.1030927835
500	69.0721649485
400	80.412371134
300	72.1649484536
200	73.1958762887
100	80.412371134
90	83.5051546392
80	86.5979381443
70	87.6288659794
60	89.6907216495
50	90.7216494845
40	93.8144329897
30	97.9381443299
20	98.9690721649
10	94.8453608247

Image size: (400, 200)
K=5
bins=[200, 100, 90, 80, 70, 60, 50, 40, 30, 20, 10]

Bin	Accuracy
200	69.0721649485
100	77.3195876289
90	78.3505154639
80	78.3505154639
70	80.412371134
60	84.5360824742
50	87.6288659794
40	90.7216494845
30	95.8762886598
20	96.9072164948
10	95.8762886598

Image size: (200, 100)
K=3
bins=[200, 100, 90, 80, 70, 60, 50, 40, 30, 20, 10]

Bin	Accuracy
200	73.1958762887
100	80.412371134
90	83.5051546392
80	86.5979381443
70	87.6288659794
60	89.6907216495
50	90.7216494845
40	93.8144329897
30	97.9381443299
20	98.9690721649
10	94.8453608247

Usage

• Go to src/ directory
• python3 main.py

Note

• Enable debug mode to know wrong prediction

  In src/signature_validation.py set LOG_DEBUG_ENABLE = True

Requirements

• Python 3.6+
• OpenCV 3.2
• Numpy
• Scikit-learn

Reference

https://github.com/vadi95/Signature-Verification.git

https://github.com/guilhermefloriani/signature-recognition.git

https://github.com/jadevaibhav/Signature-verification-using-deep-learning

https://github.com/luizgh/sigver_wiwd

https://github.com/Aftaab99/OfflineSignatureVerification

https://github.com/beyhangl/Signature_Recognition_DeepLearning

https://github.com/guilhermefloriani/signature-recognition