WELCOME WELCOME
Nov 19, 2014
WELCOMEWELCOME
A Neural Network For A Neural Network For Handwritten Digit Handwritten Digit
RecognitionRecognitionS.Praveen KumarS.Praveen Kumar(07Q71A0597) (07Q71A0597)
&&V.Tharun ReddyV.Tharun Reddy(07Q71A05B8)(07Q71A05B8)
Final YearFinal YearDept. of Computer Science & Dept. of Computer Science &
EngineeringEngineering
Avanthi Engineering CollegeAvanthi Engineering College
AGENDA AGENDA How and Where Handwritten Digit Recognition How and Where Handwritten Digit Recognition
is used?is used?
What are the several techniques used What are the several techniques used for Preprocessing thethe H Handwritten Digit andwritten Digit Recognition, as well as a number of ways in Recognition, as well as a number of ways in which neural networks were used for the which neural networks were used for the recognition task?recognition task?
Main goal of this system?Main goal of this system?
IINTRODUCTIONNTRODUCTION
A handwritten digit recognition system was used to A handwritten digit recognition system was used to visualize artificial neural networks. It is already widely used visualize artificial neural networks. It is already widely used in the automatic processing of bank cheques, postal in the automatic processing of bank cheques, postal addresses, in mobile phones etcaddresses, in mobile phones etc
To perform digit recognition, some basic knowledge on To perform digit recognition, some basic knowledge on neural network and image processing is needed. But, the neural network and image processing is needed. But, the customer may use it without any prior knowledge in image customer may use it without any prior knowledge in image processing or neural network.processing or neural network.
Some of the existing systems include computational Some of the existing systems include computational intelligence techniques such as artificial neural networks or intelligence techniques such as artificial neural networks or fuzzy logic, whereas others may just be large lookup tables fuzzy logic, whereas others may just be large lookup tables that contain possible realizations of handwritten digits.that contain possible realizations of handwritten digits.
Neural NetworksNeural Networks
Artificial neural networks have been developed since the 1940s.Artificial neural networks have been developed since the 1940s.
Artificial neural networks, usually called neural networks (NNs), Artificial neural networks, usually called neural networks (NNs), are systems composed of many simple processing elements are systems composed of many simple processing elements (neurons) operating in parallel whose function is determined by(neurons) operating in parallel whose function is determined by
1) Network Structure1) Network Structure
2) Connection Strengths and2) Connection Strengths and
3) The Processing performed at Computing 3) The Processing performed at Computing elements or nodes. elements or nodes.
TWO MAIN GROUPSOF
NEURAL NETWORK
SUPERVISEDLEARNING
UN-SUPERVISEDLEARNING
Handwritten DigitHandwritten DigitRecognitionRecognition
Preprocessing:Preprocessing: It is used to eliminate extra problems due to non-class-specific differences like
1) Size2) Shear3) Line thickness4) Background and Digit colors5) Resolution, etc.
FEW TECHNIQUES TOFEW TECHNIQUES TOMINIMIZE THE DIFFERENCESMINIMIZE THE DIFFERENCES
a) Deskewing:
The deskewing process works as follows:
1.An imaginary line is drawn through the digit:This imaginary line can be described by an equation x = ay + b, where x and y are the horizontal and vertical pixel index, respectively, and a and b are line parameters.
2.A side slip is performed:In formula: xnew = xold − ay for every pixel, resulting in the new line x =b.
3.Slant is removed by performing rotation:
(d) right-leaning (e) side slip (f) rotation
(c) rotation(a) left-leaning (b) side slip
b) b) Thinning:Thinning:
(a) threshold original (b) thinned (with spurs) (c) Pruned (by 5 pixels)
The digit images are thinned to a thickness of one pixel.
This can lead undesired side-effects ,such as spurs.
The benefits of the thinning process far outweigh the negative sides.
(d) smoothed (e) thinned
The need for pruning can be avoided by Smoothing the handwritten digit image before thinning, as seen in Figs. d and e.
c) c) SmoothingSmoothing::
Some acquisition methods may cause Some acquisition methods may cause small artifacts on the digit’s edges. small artifacts on the digit’s edges. These are removed by a smoothing These are removed by a smoothing operation.operation.
Effective smoothing can be Effective smoothing can be accomplished by filling tiny gaps in both accomplished by filling tiny gaps in both the foreground and the background.the foreground and the background.
d) d) Resizing:Resizing:
Differences in digit size are due to different Differences in digit size are due to different image resolutions and handwriting styles.image resolutions and handwriting styles.
To minimize the effect of these To minimize the effect of these differences, all digits are differences, all digits are resampled/rescaled to the same size in resampled/rescaled to the same size in pixels, before being fed to the recognizer.pixels, before being fed to the recognizer.
RESULTSRESULTS
A neural network, which is trained with 10000 handwritten A neural network, which is trained with 10000 handwritten digit images from in image database, is shipped together digit images from in image database, is shipped together with thiswith thislibrary as default neural network model.library as default neural network model.
By performing benchmarking on 8920 handwritten digit By performing benchmarking on 8920 handwritten digit images from image database, the following recognition images from image database, the following recognition raterateis achieved.is achieved.
RECOGNITION RATING TABLERECOGNITION RATING TABLE
SET OF HANDWRITTEN SET OF HANDWRITTEN DIGITSDIGITS
a) a) TRAININGTRAINING b) b) TESTINGTESTING
Example for a single Digit Example for a single Digit RecognitionRecognition
Example for a String RecognitionExample for a String Recognition
IPHONE-MOBILESIPHONE-MOBILESWITH HANDWRITTEN DIGIT WITH HANDWRITTEN DIGIT
RECOGNITION FEATURERECOGNITION FEATURE
BEFOREBEFORE AFTERAFTER
CONCLUSIONCONCLUSION
The main goal is purely educational one, a moderate recognition rate of 98% was reached on a test set.
The The Handwritten Digit RecognitionHandwritten Digit Recognition is used is used to recognize the Digits which are written by hand.to recognize the Digits which are written by hand.
The handwritten digit recognition have also proven to be a The handwritten digit recognition have also proven to be a good neural network architecture and application for the good neural network architecture and application for the purpose of introducing and demonstrating neural networks purpose of introducing and demonstrating neural networks to the general public.to the general public.
ReferencesReferences
1) Kohonen, T.: Self-Organizing Maps. 1) Kohonen, T.: Self-Organizing Maps. Springer- Verlag, Berlin (1995)Springer- Verlag, Berlin (1995)
2) Berend-Jan van der Zwaag Euregio - 2) Berend-Jan van der Zwaag Euregio - Computational Intelligence Center - Dept. Computational Intelligence Center - Dept. of Electrical Engineering - University of of Electrical Engineering - University of Twente Enschede, the NetherlandsTwente Enschede, the Netherlands
ANYANYQUERIES ?QUERIES ?
THANKTHANKYOUYOU