Natural image classification in nonlinear network model
Hanif Azhar, Khan M. Iftekharuddin, Róbert Kozma, Abhinav Admala · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2005
We study the non-linear behavior of the KIII model for natural image classification. The KIII model is designed to be a dynamic computational model that simulates the sensory cortex. The KIII model has been explored for rudimentary pattern recognition and classification in noisy environment. We extend the study of KIII models in understanding whether self-organized neural populations can be exploited into perceptual and memory producing systems such as in natural image classification. Our goal is to obtain a quantitative index on how well the KIII model behaves when it is assigned the task to identify and distinguish one class of natural image from the other based on color and texture features. For twenty training data, twenty validation data and eighty test data set for four image classes, we obtain 80% correct classification using the KIII. We compare a standard non linear neural network tools such as back propagation for the classification of the same set of natural images and obtain 65% correct classification. We conclude that dynamic neural computational models such as KIII may be suitable candidates for natural image classification.