Image segmentation using self-development neural network-applied to active stereo vision

We develop a self-development neural network (SDNN) useful in performing image segmentation. SDNN is successfully applied to improve performance of our previous work where an active stereo vision system was built. Each neuron in SDNN is characterized by a measure of vitality. By utilizing the vitality conservation principle, we show that SDNN achieves biologically plausible vector quantization, as well as facilitating systematic derivations of learning parameters. The segmentation results obtained by SDNN can serve as important cues to effectively separate objects but also help obtain the accurate outline of each object. The segmented results enables the system to quickly adjust camera positions to the chosen object, and to obtain an accurate range map as well.