Images Classification by Pulse Coupled Neural Networks

Rafidison Maminiaina Alphonse, Ramafiarisona Hajasoa Malalatiana

Abstract: The purpose of this paper is a presentation of new method of images classification. When we talk about this subject, the first reflex is thinking on convolutional neural network (CNN) such as LeNet, AlexNet, GoogLeNet, ResNet, etc. They have a good performance however another way to improve always exists. We introduce the notion of foveation which consists of collecting all pertinent information in different region of an image. Pulse coupled neural networks (PCNN) is a strong tool to accomplish this foveation task. Once, essential information is extracted, we cannot forward directly to fully connected neural network (FCNN) due of large data quantity so we compress them with Haar wavelet transform. Reshape compressed picture will be presented to FC. This neural network ensures the images classification as per the input. The singularity of this approach is the minimum time response and high accuracy percentage. Output’s total value is one because softmax function is the activation function for last layer. The neuron which has higher value indicates the corresponding class of the image.

Keywords: blurring filter, foveation, pulse coupled neural networks, wavelet transform, fully connected neural network, softmax, accuracy