Training with low number of samples

[sinangokce]

Hello,

I read your paper Segmentation-Based Deep-Learning Approach for Surface-Defect Detection and amazed by your results. However, I couldn't understand very well what property of your network allows a training with low number of samples regarding only the segmentation part. I would be very happy if you could give an explanation.

Best regards!

[skokec]

Hi,

Sorry for late reply.

The problem with limited sample sizes is that due to a large number of parameters (over 1 - 10 mio) the problem is overdetermined and model will overfit those parameters to the small training set.

But with segmentation, each pixel (or group of around 8x8 pixels, depending on down-sampling) is its own sample. This means that even in one high-resolution image you would get over 1000 or 10.000 training samples, which would limit the overfitting issue. In our network, then just by having lower layer (with most number of parameters) trained first on segmentation will result in good features for classification, which then allows to learn top features for classification from only a limited samples size.

I hope this explains it.

Best, Domen