Facial Recognition with Neural Networks Matt Fig Image Processing Fall, 2013 Introduction to Neural Networks • Attempts to imitate human neurology • Composed of nonlinear computational functions called neurons • Requires training with known data sets • For linearly separable data sets, convergence is guaranteed • More complex data sets use a technique called back propagation Introduction Continued • Consider two sets, each with two points given by ω1={(0,0,1),(0,1,1)} and ω2= {(1,0,1),(1,1,1)} • Want to classify unknown point • Algorithm finds “decision boundary” between given sets • In this case the boundary is a line • Minimize D D y w y T Example • To do this we iterate according to (for some positive constant c) wk cyk if yk 1 and wTk yk 0 wk 1 wk cyk if yk 2 and wTk yk 0 wk 1 wk otherwise Example • If we choose w1 0,0,0 T • The algorithm converges to T w 2,0,1 in 14 iterations. Example Continued • Equation of line is D y 2 x1 1 0 x1 1 2 Advanced Networks • Need something else for non-separable data sets • Here a minimization function of this form is used 2 1 T J w r w y 2 • r is the desired response for the particular training vector y • Minimizes least squared error Advanced Networks • Weight vector becomes wk 1 wk rk wTk yk yk • To establish a network, create several layers of these neurons connected such that the output of every function in one layer is an input into the functions in the next Neural Network Facial Recognition • Many MATLAB solutions online • One was chosen as a basis for project • Trained with 64 random faces Testing • The trained network was tested for effectiveness as a function of noise and lighting intensity • 5 people were used for experiments • Gaussian noise was added with varying intensities • Light was adjusted 5 stops with dimmer • As expected, more noise leads to more misses in identification, same with lighting. Gaussian Noise Dimming Re-training • After these results, noisy and dimmed images were added to the training set. • This results in across the board improvement in identification success • For a complete and reliable system, many more faces in many more situations should be added to the training database Questions?