Real time object recognition for teaching neural networks
F. Khan, Alejandro Cervantes · 2003
Undergraduate students in computer science learn best when they are given the opportunity to apply hardware and software concepts to real world systems, and neural-network applications present attractive possibilities for giving them such opportunity. An example of how to take advantage of these possibilities is given in this paper, which describes a specific neural network technique that has been developed and applied to the problem of identifying real world objects in real time. Those objects can be as simple as paper cut-outs or they can be mechanical objects such as a nut or a bolt. These objects are placed on a plane, and are "examined" by an "identifying system" consisting of a camera attached to a PC through a video capture card. The pixels collected from the image of the object are fed to a set of neural network nodes for pattern recognition. Patterns are recognized by a multi-layer neural network where the output of a neuron, which can be characterized by sigmoid "activation function", is a function of weighted inputs. Lisp is used as the programming language due to its simple syntax and powerful recursive features for processing lists. The binary equivalent of the computed output is evaluated as a recognition signature which is compared to the signature of objects in a list. Of course, the artificial neural network is "trainable". To master the technique, the students start by learning about neurons and forward and backpropagation methods, but they soon find themselves "training" a multilayered neural network that they themselves have built. The learning experience encompasses video capturing, image handling, filtering, and image compression, and it demystifies neural network programming.