Neural network problems with multi-input dependencies

[roldy]

I'm working on a fingerprint recognition project and have run into a road block. As of now, I am using a 3 layer feed-forward neural network to find the minutiae in a fingerprint. If a minutiae is detected, it's location in the image (row, col) and the type (n types) of bifurcation is noted; these parameters will be known as properties. I do this for my target (database) set and for my input fingerprint. What I want to do now is to maybe use another neural network to compare the properties of the input fingerprint to the properties of the fingerprints in the database.

The problem here is that I basically have 3 property values. row can range from 1 to max row in target and col can range from 1 to max column in target. The type n, for the sake of explanation, is say 6. Is it possible to train a neural network to recognize any integer value for these three properties?

Another possibility that I was thinking of implementing was the use of an error function. I could come up with some sort of error function that's dependent on the input fingerprint and the target fingerprint properties. Basically the error function would return a 1 or 0 depending on if the value falls or exceeds a certain threshold. The problem with this is that I would have different threshold values.

Any thoughts or suggestions on the feasibility or possibility of such a neural network/error function? Your help is greatly appreciated.

 

[lofgran]

I understand your frustration with running into roadblocks in your project. It seems like you have already made a lot of progress with using a 3 layer feed-forward neural network to detect minutiae in fingerprints. However, I can see how the issue of having multiple property values for each fingerprint could be a challenge.

To answer your first question, yes, it is possible to train a neural network to recognize any integer value for these three properties. This is because neural networks are highly flexible and can be trained to recognize patterns and relationships between different variables. However, it may require some additional preprocessing of your data to make it more manageable for the neural network to handle. For example, you could try converting the row and column values into relative positions within the fingerprint image, rather than absolute values.

As for your second idea of using an error function, it could also be a viable approach. However, as you mentioned, the challenge would be determining the appropriate threshold values for each property. This could potentially require some trial and error, and it may not be as accurate as using a neural network.

In terms of suggestions, have you considered using a convolutional neural network (CNN) for your project? CNNs are specifically designed for image recognition tasks and may be better suited for your fingerprint recognition project. Additionally, you could also try using a combination of different neural networks, such as using a CNN for feature extraction and then feeding those features into a feed-forward neural network for classification.

Overall, I think both of your ideas have potential and it may be worth exploring both options to see which one yields better results. It's also important to continue gathering and analyzing data to ensure that your neural network is being trained on a diverse and representative dataset. I wish you the best of luck with your project and hope that my suggestions are helpful.