In number classification using neural network, is training with edge image better than gray image?

Question

i'm trying to identify numbers and letters in license plate. License plate images are taken at different lighting condtion and converted to gray image. My concern with type of data for training is:

Gray Image:

• Since they are taken at different lighthing condition, gray image have different pixel intensity for same number. Which means, i have to get many training data for different lighting condition to train.

Edge Image:

• They lack enough pixel information since only edge is white while others(background) are black. So i think they will be very weak for translational difference like shearing or shifting.

I want to get some information about which type of image is better for training number in different lighting condition. I wish to use edge image if they don't differ much since i can prepare edge image right now.

Answer 1

Theoretically you will have no gain in the error ratio if the system preprocess the images with a linear high-pass filter before to send the image to the NN.

Let see a simple 1-dimension case that supports this statement:

Assume inputs are "a", "b" and "c". A node at first hidden layer will receive an input to the activation function equal to s=w1*a+w2*b*w3*c+... being w1, w2, w3, ... the weights for this node.

Now, assume a simple differential case, where inputs will be ... , b'=b-a, c'=c-b, ... being the input to the hidden node s=w1'a'+w2'b'+w3'c'+...=...+w2'(b-a)+w3'(c-b)+...=...+(w2'-w3')b+...

Note both inputs to hidden node are the same if w1=w1'-w2', w2=w2'-w3', ... . So, the NN can itself perform the equivalent of the linear high-pass filtering adjusting the weights of the first hidden layer.

However, changes in learning speed (convergence time) can be expected.

(ps: please, activate tex/latex in this site, urgently ! )