This article describes the structure of the Image Receptive Field Neural Network (IRF-NN), introduced recently by our team. This structure extends simplified learning introduced by Extreme Learning Machine and Reservoir Computing techniques to the field of images. Neurons are organized in a single hidden layer feedforward network architecture with an original organization of the network's input weights. To represent color images efficiently, without prior feature extraction, the weight values linked to a neuron are determined by a 2-D Gaussian function. The activation of a neuron by an image presents the properties of a nonlinear localized receptive field, parameterized with a small number of degrees of freedom. This article shows that an efficient representation of the images is provided by a large number of neurons, each associated to a randomly initialized and constant receptive field. The training step determines only the output weights of the network. It is therefore extremely fast, without retropropagation or iterations, and remains efficient with large sets of images. The network is easy to implement, presents excellent generalization performances for classification applications, and allows the detection of unknown inputs. The efficiency of this technique is illustrated with several benchmarks, photo and video datasets.

No abstract