Déjà Vu object localization using IRF neural networks properties

This article introduces an original method of image detection and localization in a picture by scoring the output of a neural network to indicate an already seen (Déjà Vu) input. The classifier is a feedforward multi-layer perceptron adapted to supervised image recognition named Image Receptive Fields Neural Network (IRF-NN). It has interesting properties: fast and efficient training, as well as accurate classification skills on large learning sets. We show that a simple analysis of the neural response can be used to evaluate the probability that an input is known. This evaluation can be efficiently used to detect and localize objects in a picture using a sliding window approach. The generalization skills of the IRF-NN induce nice properties that improve the recognition process and the speed of the algorithm.