| Recognizing Rotated and Occluded Shapes Using Cortronic Neural Networks and Biologically Motivated Feature Extractors (1999) | |||||||||||||
Abstract | |||||||||||||
| A new object recognition system that uses biologically motivated feature extractors and the cortronic neural network is presented. Images are processed by edge detectors, orientation selective units and end-stop cells, in analogy to the on-center/off-center, simple, complex and hypercomplex cells of the mammalian retina and visual cortex. The resulting image is a sparse pattern suitable for input to the cortronic neural network which is used for classification. Three types of experiments are performed: recognizing rotated images, recognizing occluded images and detecting images in cluttered scenes. For images rotated within an 85 ◦ range, recognition accuracy is 93.3 % with 15 pattern classes, and 89.3 % with 30 classes. This system differs from other related work ([4], [8], [9]) in that it uses the recently developed cortronic neural network, which is a biologically motivated associative memory [6]. 1 | |||||||||||||
Publication details | |||||||||||||
| |||||||||||||