Goal of Project
The goal of this project was to take theoretical neural network learning and apply it to a product that abstracts the complexities of implementation to solve an actual problem. LizardNet presented this challenge. One of the LizardNet project was to identify lizards in images. This identification would be time consuming to be done by hand, and electronic records would not be kept. By automating the task in the form of a web application and automatic tool, the data can be analyzed without as much time and effort.
An example of the classification being attempted by this project is as follows:
We have an blank image of the lizard pen before the lizards were placed in the pen.
We also have an image of the current situation in the lizard pen.
Background
The LizardNet project is a joint effort between the Harvey Mudd College Biology Department and the Computer Science department. The lizard classification neural network is a class project relating the the LizardNet project. It attempts to solve one of the more laborious tasks involved with the study of lizards by automatically locating lizards in images of outdoor terrain.
Process
A neural net is used to train each positive image. Then negative images are fed into the network, and the neural network is trained against the false positives detected. The resulting network weights are saved for later use classifying lizards.
Results
There are no results yet.
Code
There are two main code components:
Hand Classifier Java Applet
In order to simplify the process of entering data manually, a Java applet was written to communicate back to a central server with the hand classification data.
Try it out: here
Get the code: here
Neural Network Java Tool
The actual tool is written in java, and comes with preliminary starting weights derived from initial testing. The use is simple, and is explained here.
Get the code: here
References
I researched various topics involving detection of objects using machine learning before conducting this project. Many useful ideas were borrowed from Carnegie Mellon's Computer Vision lab.
Mike Buchanan's back propagated network in Java was used for the project.