The document describes a captcha cracking system that uses machine learning techniques. It segments captcha images into characters, extracts feature vectors, and uses these in a dataset to train classification models like decision trees, neural networks, SVM, and Naive Bayes. It finds that an artificial neural network achieves the best performance with 83.8% correct classification in 10-fold cross-validation testing. The system demonstrates an intelligent way to recognize captcha patterns and crack captchas online through learning.

1. Omer Shafiq FA09-BCS-098
Ihsan Ullah FA09-BCS-153
Adnan Bajwa FA09-BCS-163

2. Brief Description
• Captcha Cracking System cracks the captcha
images intelligently and then make the
knowledge-base of the policy of cracking
captcha images and reflexly learn this
knowledge to crack the another captcha
images
• Enables the system to implement the online
learning through which we can achieve the
optimal solution
• Our System depends on multiple phases,
which are explained below that how each of
them works and integrate to make this
System.

3. The Process
• The Learning Process Takes Place After Creating Instances
List From Filtered Data

4. Analysis
• CAPTCHA IMAGE: Our System will be working on
CAPTCHA breaking written in Java using some
external OCR libraries and some Machine
Learning Libraries.
• DE-NOISE: For the first section, de-noising, we
will have to find a smart way to de-noise our
input CAPTCHA via some image de-noise
algorithm for our approach.

5. Analysis
• SEMENTATION: For the segmentation stage, we need
to split the image of string in characters via different
segmentation algorithms.
• BINARY BIT STREAM: Segmentation gives us the
different segmented images.

6. Analysis
• DATASET: Data Set contains the instances includes the
feature vectors and desired target output value which
will be predict through applying desired Classifier.

7. Architecture

8. Intelligence Aspect
• Project Intelligently Recognizes The
Pattern of the image to classify
• Project can simultaneously denoise
and segment captchas parallel
• Classification depend upon the model
you have trained

9. Application Screenshots
• CAPTCHA CRAWLER Crawels as many captchas as
we want from captchas.net server created on
C#.NET Framework 4

10. Application Screenshots

11. Application Screenshots

12. Application Screenshots

13. Results and conclusion
Classifier: Decision Tree (J48)
Instances: 353
Attributes: 191
Test mode: 10-fold cross-validation
Correctly Classified Instances 168 47.7273 %
Incorrectly Classified Instances 184 52.2727 %
Kappa statistic 0.4519
Mean absolute error 0.04
Root mean squared error 0.183
Relative absolute error 54.4013 %
Root relative squared error 95.4941 %

14. Results and conclusion
Classifier: Artificial Neural-Net (MultiLayer-Preceptron)
Instances: 353
Attributes: 191
Test mode: 10-fold cross-validation
Correctly Classified Instances 295 83.8068 %
Incorrectly Classified Instances 57 16.1932 %
Kappa statistic 0.8301
Mean absolute error 0.0171
Root mean squared error 0.0966
Relative absolute error 23.2233 %
Root relative squared error 50.4266 %

15. Results and conclusion
Classifier: Support Vector Machine(SVM)
Instances: 353
Attributes: 191
Test mode: 10-fold cross-validation
Correctly Classified Instances 304 86.3636 %
Incorrectly Classified Instances 48 13.6364 %
Kappa statistic 0.8569
Mean absolute error 0.0711
Root mean squared error 0.1861
Relative absolute error 96.8087 %
Root relative squared error 97.1338 %

16. Results and conclusion
Classifier: Naive Bayes (NaiveBayesin)
Instances: 353
Attributes: 191
Test mode: 10-fold cross-validation
Correctly Classified Instances 268 76.1364 %
Incorrectly Classified Instances 84 23.8636 %
Kappa statistic 0.7499
Mean absolute error 0.018
Root mean squared error 0.1282
Relative absolute error 24.5384 %
Root relative squared error 66.9059 %

17. Visual Results and Conclusion
Results and conclusion
100
90
80
70
60
50 Correct Classification
Missclassification
40
30
20
10
0
Decision Tree SVM ANN NaïveBayes