Image Categorisation using Feed-Forward Network. Why do i get the same values?

Question

I'm pretty new to all this business of machine learning and image categorisation. I'm using pybrain as my primary neural network modelisation tool.

My problem is the following. I have 100x100 images that have 0,1,2,3 items and I would like the machine to tell me how much elements are in a new 100x100 image.

I made a training set of 25 elements (just to see if I could get something out of that). I realise that this number is very very small relatively to the required test set but gathering training data would be very expensive (doable but expensive).

Here is how i made the training set

    onlyfiles = [f for f in filter(lambda x: x[-3:] == "png", onlyfiles)]
    input = np.zeros(shape=(len(onlyfiles), 10000))
    output = np.zeros(shape=(len(onlyfiles), 1))

    row = 0
    for file in onlyfiles:
        img = Image.open(path+file)
        pix = img.load()
        for x in range(100):
            for y in range(100):
                input[row, 100*x+y] = pix[x, y][2]
        output[row,0] = output_mapping[file]
        row += 1

I only used the pix[x,y][2] (blue color) because the zone is lightened by blue light.

So I generated the classification data set

    ds = ClassificationDataSet(10000,1,nb_classes=4)
    for k in range(len(X)):
        ds.addSample(np.ravel(input[k]),output[k])

    temp_testdata, temp_traindata = ds.splitWithProportion(0.25)

    tstdata = ClassificationDataSet(10000,1,nb_classes=4)
    trndata = ClassificationDataSet(10000,1,nb_classes=4)

    for k in range(len(temp_testdata)):
        tstdata.addSample(temp_testdata.getSample(k)[0],temp_testdata.getSample(k)[1])

    for k in range(len(temp_traindata)):
        trndata.addSample(temp_traindata.getSample(k)[0],temp_traindata.getSample(k)[1])

created the network

    fnn = buildNetwork(trndata.indim, 1, trndata.outdim, outclass=SoftmaxLayer)
    trainer = BackpropTrainer(fnn,dataset=trndata, momentum=0.1, learningrate=0.01, verbose= True, weightdecay=0.01)

trained the network

    trainer.trainEpochs (50)

and tried to push some data back in the

    fnn.activateOnDataset(tstdata)

and here it is what I got

    Out[186]: 
    array([[ 0.20086945,  0.46216524,  0.30707544,  0.02988986],
           [ 0.20086945,  0.46216524,  0.30707544,  0.02988986],
           [ 0.20086945,  0.46216524,  0.30707544,  0.02988986],
           [ 0.20086945,  0.46216524,  0.30707544,  0.02988986],
           [ 0.20086945,  0.46216524,  0.30707544,  0.02988986],
           [ 0.20086945,  0.46216524,  0.30707544,  0.02988986]])

From that I understand that my neural network would give Class "2" for all tstdata samples. I'm surprised. I was expecting crappy results, but I did not expect the very same 6 outputs...

And for the sake of checking out, here it is the "sum" of all pixel for each of those 5 samples showing that my 5 samples are indeed different...

    In[199]: np.sum(tstdata.data['input'],1)
    Out[198]: array([ 1170985.,  1115688.,  1173766.,  1125111.,  1152222.,  1132945.])

I even had thrown bunches of 0 and 1 as "fake" sample and the numbers did not change!

Any idea why all my trails have the very same output?

Thanks a lot!