|
On Aug 25, 6:39 am, "Kueh Lydia" <leehk...@naver.com> wrote:
> Hi,
>
> I am doing face recognition with PCA and back propagation neural network toolbox too.
>
backpropagation is the name of a type of optimization algorithm
that is used on many different topologies of neural networks
it is neither the name of a network nor the name of a toolbox
> I read your asking help inhttp://www.mathworks.fr/matlabcentral/newsreader/view_thread/302393.
I did not ask for help.
> Did you manage to implement the recognition task??
I have given advice on pretraining and code in many posts.
For advice, search the archives using
heath pretraining advice
heath face recognition
> I am writting to ask help too.
> Please respond as soon as possible.
What image size?
What range of pixel values?
Converted to what vector length I0 = ?(eye-oh, not ten)
> I have calculated the eigenfaces
Using what eigenvalue threshold?
size(p0) = [I0 N] > PCA => size(p) = [I N]
>but i dont know how to implement recognition task with neural network toolbox too.
>
> My database is 175 images from 35 person (each person 5 images)
>
> The out put of PCA algorithm is a weight matrix B of 103 *175
> use it as the input of neural network.
delete the word "weight"
I doubt if you need that many input variables. Review your
eigenvalue threshold.
Original imput matrix
size(p0) = [I0 N] = [ I0 175]
PCA Transformation matrix
size(T) = [ I I0] = [ 103 I0 ]
Transformed input matrix
p = T*p0
size(p) = [ I N ] = [ 103 175]
Ourtput matrix
size(t) = [O N ] = [ 35 175 ] % Columns of eye(35)
> The weight matrix means 1 image with 103 features. 175 for
>35 person of 5 images
Standardize (zero mean/unity variance) inputs
Training will yield Neq = N*O equations for Nw weights
Design a constant naive matrix model with constant outputs equal
to the mean of the targets forming Nw00 = O weights. Obtain
MSE00 = SSE00/Neq, the degee of freedom adjusted value
MSE00a =SSE00/(O*(N - Nw0)) and corresponding R^2 statistics.
(Also classification error rates for each person)
Design a linear matrix model with Nw0 weights. Obtain MSE0,
MSE0a, corresponding R^2 statistics with R20 = 1-MSE0/MSE00
and R20a = 1-MSE0a/MSE00a = 1-(N-Nw00)*(MSE0/MSE00)/(N-Nw0)
and classification error rates for each person
> I have created the nn with newff and train the network too.
> net = newff(minmax(B),[S1 S3],{'logsig' 'logsig'},'traingdm');
Single hidden layer node topology I-H-O
Try to choose H so that there are many more training
equations than unknown weights with the constraint
Neq = N*O >> Nw = (I+1)*H+H*(I+1) ==> Hmax
Choose R2agoal (e.g., 0.99)
Choose or loop over 1 <= H <= Hmax
net = newff(minmax(p),[H O],{'tansig' 'logsig'});
Nw = O + (I+O+1)*H
MSEgoal = (1-R2agoal)*(Neq - Nw)*MSE00/(N - Nw00)
net.Parameter.goal = MSEgoal;
net.Parameter.show = 10;
[net tr Y E] = train(p,t);
Nepochs = tr.epochs(end);
MSE = tr.perf(end);
R2 = 1-MSE/MSE00
MSEa = N*MSE/(N-Nw)
> The target is an identity matrix T of 175 *175.
No. See above for size(t)
How to set the target T??
> [netl,trl, trlO, trlE] = train(net,B,T);
>
> My testting database, BI1 is 175 images from 35 person
(each person 5 images) too. They are different sets of images
from 35 person.
>
> I simulate using A=sim(net,BI1);
> Follow by command AA = compet(A).
> Then I calculated the error by errors1_1 = errors1_1 + sum(sum(abs(AA-T)))/2;
>
> Does it correct?
>
> Actually, I I divided my database as follow
>
> Training set Testing set
> 5 images per person 5 images per person
>
> 4 images per person 6 images per person
>
> 3 images per person 7 images per person
>
> 2 images per person 8 images per person
>
> 1 images per person 9 images per person
>
> But I don't know how to simulate the train net with the testing
> images from these 35 person of my database.
>
> The target matrix for testing set and training set are not the same.
The lengths are different. They are not match each other.
>
> How to simulate them and how to count the error of testing set??
>
> Could you please help me on this part. It is urgent
1. Normalize the test set ptst using the means and standard
deviations
of the unstandardized training set
2. ytst = sim(net,ptstn)
3. etst = (ttst-ttst),
4. MSEtst = mse(etst)
5. R2tst = 1-MSEtst/MSEtst00
Hope this helps
Greg
|
Feed for this Thread