Outline

• Problem Statement and Objective
• Introduction to the Domain - Machine Learning and Deep Learning
• Literature Survey
• Requirements
• Analysis - Existing VS Proposed

Outline

• System Architecture
• Project Features
• Algorithms
• Implementation
• Demo
• Benchmark Dataset

Outline

• Performance Metrics
• Test Case
• Findings
• Future Work
• References

Problem Statement

Problem Statement

Image Understanding Systems can help us a lot

Eg. Cheques, Number Plates, Passports

Objectives

• To develop a system that can classify tamil characters using raw pixels
• To minimise the amount of human assistance required for the system
• To understand to how much extent a linear classifier can be used on the dataset
• To understand to how much extent, the architecture of a neural network affects the classification rate
• To understand how a deep learning algorithm performs on the dataset
• To learn machine learning and deep learning

Introduction to the Domain of Work

Machine Learning

In Simpler Terms, Decision Making without if-else statements

Motivation

This is what the computer sees

Possible Solutions

Literature Survey

On the Previous Works in Tamil Character Classifiaction

• Two stage process
• Dataset used : HPL Tamil
• First Stage : Classes are grouped into clusters - K means clustering
• Same class can be in different clusters
• MLP

• Two feature extraction First stage - 7 x 7 boxes Second stage - chain code histogram
• Two stages
• Only Clusters Prediction

Literature Survey

On the Previous Works in Tamil Character Classifiaction

• Shows a system that can recognise characters without considering the translation and scale factors
• Origin is set to the center of a character
• Several network architectures are discussed

• Uses Online data
• Feature extraction - calculation of new xy coordinates

Literature Survey

On the Previous Works in Tamil Character Classifiaction

• Shows all the work done in the recognition of South Indian Scripts
• Some common pre processing techniques
• Results obtained

• Most were online data

Literature Survey

On the Previous Works in Tamil Character Classifiaction

• Uses an SVM to classify images
• feature extraction - image divided into 64 zones
• pixel densities are calculated for each zone

• manual feature extraction
• two step extraction

Literature Survey

On the Previous Works in Tamil Character Classifiaction

• feature selection through Zoning and Chain code
• feature extraction through sub line direction and bounding box algorithm
• SVM

• manual feature extraction
• two step extraction

Literature Survey

On the Previous Works in Tamil Character Classifiaction

• Review Paper
• Presents a survey of the work carried out in tamil character recognition

• All had separate feature extraction stage

Literature Survey

On CNN Architecture

• AlexNet
• Filters:
  • 11 x 11 with stride 4
  • 5 x 5
  • 3 x 3

Literature Survey

On CNN Architecture

• VGGNet or the OxfordNet
• Filters:
  • 64, 128, ...
  • 3 x 3 Fixed
  • stride: 1
  • padding: 1
• Max-Pooling:
  • 64, 128, ...
  • 2x2
  • stride: 2

Dropouts and other Hyper parameters

Online Courses

• Stanford's CS231n
• Andrew NG's Machine Learning Course

Hardware

• 16 GB RAM
• Intel i7 3rd gen processor

Software

• Python
• numpy
• scikit-learn
• PIL
• ipython
• matplotlib
• theano
• lasagne

Existing System

Manual Feature Extraction Required

Proposed System

Manual Feature Extraction NOT Required

System Architecture

System Architecture

Project Features

• Offline based - direct images
• Can support any image formats
• No hand engineered features required
• Single stage process
• Output classes are not clustered
• Limited training examples

Project Features

• 45 class classifier
• Prints tamil characters - unicode
• Works with images that are compressed
• No noise removal stage
• Open Source

Algorithms

Softmax Classifier

• Score Function : y = ${softmax}(x)_i$ Here softmax is : ${softmax}(x)_i = \frac{\exp(x_i)}{\sum_j \exp(x_j)}$ And x is the result of the linear function W(input) + b
• Initialise the weights and biases
• Loss Function : $-\log\left(\frac{e^{f_{y_i}}}{ \sum_j e^{f_j} }\right) \hspace{0.5in} \text{or equivalently} \hspace{0.5in} L_i = -f_{y_i} + \log\sum_j e^{f_j}$
• Set the hyper parameters
• Run an optimisation algorithm (SGD)

Algorithms

2 Layer Neural Network

• ReLU: max(0, x)
• Gradient Descent

Algorithms

3 Layer Neural Network

• ReLU: max(0, x)
• Gradient Descent with Nesterov Momentum

Algorithms

Convolutional Neural Network

• Weights - Xavier Initialization technique
• Gradient Descent with Nesterov Momentum

Algorithms

Convolutional Neural Network

• Non Linearity : ReLU
• Filters :
  • Size : 3 x 3
  • Number of Filters : Conv_1: 32, Conv_2: 64
  • Zero Padding: 1 pixel
• Max-Pooling :
  • Size : 2 x 2
  • Stride : 2
• Dropout is added for regularization
• Number of hidden units : FC_1 layer - 250 , FC_2 - 45

Implementation

System = API + ipython notebook client

# Now Start training by a creating a ConvNet Instance
	from ConvNet import ConvNet

	net = ConvNet()
	model = net.fit(X_tr, y_tr, X_val, y_val, X_ts, y_ts)

ConvNet ipython notebook

Implementation

Demo

Results Obtained

Linear Classifier

Results Obtained

Two Layer Neural Network

Results Obtained

Three Layer Neural Network

Results Obtained

Convolutional Neural Network

Performance Metrics

Training and Test set accuracy - mean(prediction == actual) * 100

Performance Metrics

Loss Graphs and Filter Visualisations

Findings

• Linear Classifers
  • Took the least training time
  • Produces average results
  • Not a good idea to try on images
• Adding layers to an existing neural network architecture will not drastically improve the accuracy
• Plot graphs

Findings

• Tuning Hyperparameters takes time
• The number of hidden units (hyperparameter) does not affect accuracy much ( Eg - 1000 units and 2000 units )
• Tuning hyperparameters can improve accuracy tremendously (2% - 52%)
• Mini Batch training is better

References

• [1]. Bhattacharya U, Ghosh SK, Parui SK (2007) A two stage recognition scheme for handwritten Tamil characters. In: Proceedings of the ninth international conference on document analysis and recognition (ICDAR 2007). IEEE Computer Society, Washington, DC, pp 511–515
• [2]. T.Paulpandian and V.Ganapathy Translation and Scale Invariant Recognition of Handwritten Tamil Characters Using a Hierarchical Neural Network”, 1993 IEEE
• [3]. Jomy John, Pramod K. V, Kannan Balakrishnan Handwritten Character Recognition of South Indian Scripts: A Review , National Conference on Indian Language Computing, Kochi, Feb 19-20, 2011
• [4]. Shanthi N, Duraiswami K A Novel SVM-based handwritten Tamil character recognition system. Springer; Pattern Analysis and Application. 2010; 13(2):173–80.
• [5]. S. M. Shyni, M. Antony Robert Ra, S. Abiram, Offline Tamil Handwritten Character Recognition Using Sub Line Direction and Bounding Box Techniques Apr 2015, Indian Journal of Science and Technology

References

• [6]. Antony Robert Raj M, Abirami S. A survey on Tamil handwritten character recognition using ocr techniques. The Second International Conference on Computer Science, Engineering and Applications (CCSEA). 2012; 5:115-27.
• [7]. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 25, pages 1097– 1105. Curran Associates, Inc., 2012.
• [8]. K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
• CS231n
• Coursera Course
• HPL Tamil Dataset