Neural networks and image processing for shooting caterpillars with lasers

Question

I'm a bit of an amateur farmer, and I have a precious potted cherry tomato plant. Recently, to my chagrin, I discovered that my precious plant was the victim of a scheme committed by the evil Manduka Quinkemekulata - also known as Tomato Horncherm ( http://insects.tamu.edu/images/insects/common/images/cd- 43-c-txt / cimg308.html ).

During the rout of the last worm that I saw, I thought to myself, if I were to use a webcam connected to a computer with a running program, can I use some kind of application to monitor my precious plant? These pests are highly masked and very difficult for my naive eyes.

I have seen studies using artificial neural networks (ANNs) for all kinds of things, such as facial recognition, etc., and it might be possible to find a pest using ANN.

I have a few questions, although I would like to get some suggestions.

1) Is there a rating of various ANNs in terms of how good they are classified? Are multilayer perceptrons known to be better than Hopfields? Or is this a question for which the answer is unknown?

2) Why are there several different activation functions that can be used in ANN? Sigmoid, hyperbolic tangent, step functions, etc. How to find out which function to choose?

3) If I had the image of a plant with a worm on one of the branches, I think that I could train the neural network to look for thin thin branches, lose weight quickly, and then get thin again. I have a problem, but with branches intersecting everywhere. Is there a preprocessing step that can be applied to an image to distinguish between foreground and background elements? I would like to isolate separate branches for working on the network one at a time. Is there any good conversion algorithm?

Any good pointers to pattern recognition and image processing, such as books or articles, would also be greatly appreciated.

Regards, MJ

Ponomar worms were damaged while writing this letter.

Answer 1

A good rule of thumb for machine learning: better features outperform algorithms. If you feed the original pixels of the image directly into your classifier, the results will be poor, no matter what training algorithm you use. If you pre-process the image and extract functions that strongly correlate with the "presence of the caterpillar," then most algorithms will do a decent job.

Therefore, do not focus on network topology; start with the task of seeing a computer.

Answer 2

Do these little suckers move regularly? If so, and if the plant is fairly static (which means there is no wind or other forces that make it move), then a simple filter is enough to search for movement. This will do without any learning algorithms, which are often quite difficult to train and implement.