neural network image classification (good, so-so, bad)

I would build a neural network to be feeded with images taken by my cell phone or camera. Starting from that image on some tools, ANN should identify 3 classes: OK, not so good, BAD.

Can someone address me on ho extract correct features from images and the feed my ANN? Any idea on how many layers and nodes per layers? I think I should have same neurons as input as pixel of my image, and 3 neurons as output.

TIA Francesco

[Greg Heath]

The number of hidden nodes, H, should be the smallest number that will yield acceptable results. I typically use Ntrials = 10 or 20 nets for each value of H in an interval Hmin:dH:Hmax.

For most of the MATLAB examples

 help nndatasets
 doc nndatasets

I have used numel(Hmin:dH:Hmax) = 10 which may have to be refined with another run with larger Hmin and smaller Hmax.

Hope this helps.

Greg

For examples search BOTH the NEWSGROUP and ANSWERS using

 greg Hmin:dH:Hmax

 greg Ntrials

Answer 1

[Explorer]

Neural Network images classification works fine. Accuracy of classification also depends on the features of the images. You haven't yet extracted features of images. Without features, you can not proceed.

First find an algorithm from research papers to extract features of images. And then you can ask for help in classification or feature matching using ANN.

By the way, Greg Heath has good knowledge of Neural Networks.

[Francesco Piantedosi]

Thank you very much for your answer! I'll visit pages on Mr. Heath , and maybe I'll ask his support in future. Great clarification in any case. My problem is about drill bits, I work for Aeronautics company and I'm focused on maintenance for handtools. I need to have an automatic classification and understand if a drill bit is re-sharpening, completely new or to be scrapped. My first attempt will be, starting from some photos, understand how I can extract features that give me these idnications. Edges worn, reduced dimensions and so on, then my next step will be identify correct ANN topology. I'll use several pictures to have my training set, validation set and Test set. I'll let you know about my progress. Maybe in the future a best approach could be start from a video, with rotating drill bits...but it's enough for me starting from some pictures and see what happens.

Thank you Francesco

[Explorer]

You welcome :)

[Francesco Piantedosi]

I summarize my results: starting from an Image, with pictures of drill bit, I crop image in order to have less pixel to analyze. Let's assume that I have X * Y pixels. Then I convert image into gray scale, so I reduce dimension of my input to my ANN,because starting from a matrix X*Y*3 I will have a matrix X*Y.

Then I move every column under the first one so I will have a single column with X*Y elements that I convert into Double.

Then my question: due to high number of X*Y, can I have a so big number of neurons as first layer?

Or do I need to collect my input data in sub-vectors to feed a single neuron with a vector?

O r my I chose a deep learning strategy in order to train different layers with different portions of my original vectors?

Of course I'll repeat this cycle for every image of drill bit, I hope 2000 pictures, so I will be able to train my network with 2000 of columns.

75% for trainig, 15% for validation and 15% for testing.

Can you help me to identify correct topology for my network, transfer function and training strategy?

TIA Francesco