Final Result

Finally we could complete the project. The results were no that great. 67% of the faces could be detected with good number of false positives. The most important thing is that we could understand the difficulty in constructing the database. We did not expect database creation to be that cumbersome. So, we could understand the practical hurdles in solving a real-world problem using machine learning methods.

Today's update

Today we studied how to use the neural network toolbox. We tried with different training data. We experimented with four(20*20) face images and soon realized that 150 images that we have is not at all sufficient for training/validation/testing. So, we again started to collect training data. Till now we are working only on (20*20) image. Next we plan on working 1. On images of all sizes 2. Create our own code for neural networks(instead of using toolbox);

At last Database is ready

Since 2-3 days, Srikanth Parupati and I have been working on creating a database of face/non-face images(of 20*20 size). The database consists of ~200 images. Now we are learning to use the Neural Network Toolbox provided by Matlab.

My Neural Network Algorithm

Since 3 days I was trying to write a sample neural network algorithm for boolean value inputs and output. My neural network contains three inputs and one output.I used one hidden layer with two hidden units. The AND operation on the three inputs gives the output. I wrote down a truth table and supplied this as a set of training examples to the algorithm. I initialized the weights to some random numbers(between 0 - 0.05). Then I ran the algorithm. After 50 iterations over the training examples, 6 out of 8 examples are consistent with the output. After that I iterated through the examples many times(nearly 2000000) but still only those 6 same samples are consistent. The input weights almost converge after 2000 trails but the hidden weights(weights on the links between hidden units and output units)never converge. They keep oscillating between 2 samples. I even tried by reducing the 'eta' value. But all 8 samples are never consistent with a set of weights.

Another good paper

There is another paper that I came across,

A HIERARCHICAL NEURAL NETWORK FOR HUMAN FACE DETECTION - PAUL JUELL and RON MARSH.

The approach used in the paper uses four neural network. Three of them are used to detect eyes, nose and mouth. The output of these networks are fed into the fourth neural network. I feel the concept and results are well explained but the implementation details are not so clear.

Interesting paper

Today I came across many good papers on face detection using neural networks. I read one of them which is very interesting. The URL is

http://www.ri.cmu.edu/pub_files/pub1/rowley_henry_1998_1/rowley_henry_1998_1.pdf

The paper is old but explains all concepts very well. The paper compares its approach with all existing approaches.

Aim of the Project

The basic aim of this project is to create a neural network that takes an image as input and tells us if the input image contains faces or not. We plan to use Backpropagation Algorithm for training. Once we complete this part we plan to extend our project by analysing different neural network algorithms and find which is most efficient.