This project develops a framework
called Cresceptron for automatic learning for
recognition and segmentation of real-world 3D objects from
their images based on exemplars of performance of such tasks.
The Cresceptron has been tested on the task of visual recognition: recognizing
3-D general objects from 2-D electro-optical images of natural scenes
and segmenting the recognized objects from their cluttered image background without handcrafting a 3Dobject model.
Specifically, it recognizes and segments image patterns that are
similar to those learned, using a stochastic distortion model and
view-based interpolation, allowing other view points that are moderately
different from those used in learning.
It incorporates both individual learning and class learning; with the
former, each training example is treated as a different individual and with
the later, each example is a sample of a class.
Several types of network mechanisms have been developed, and their properties
are addressed in terms of knowledge recallability, positional invariance,
generalization power, discrimination power and space complexity.
Experiments with a variety of real-world images
are reported to demonstrate the feasibility of the Cresceptron with an almost perfect performance.
Borrowing the three main ideas (convolution, paired layers, and increasing receptive fields from early to later layers in a deep cascade) of Neocognitron by Fukushima (which was designed for a single isolated 2D character), Cresceptron has made the following seven first-timers: