Fast Image Recognition Using Double Hyper Basis Function Neural Network and Its Combined Learning

The proposed neural network with kernel activation functions is designed to solve the image recognition-classification problem in online mode. The modified probabilistic neural network (PNN) as an encoder gets on its input layer image data, represented in matrix form. The outputs of the PNN, a number of which is defined by the amount of classes in a dataset, a vector of signals of pre-classification appears. This signal is fed to the input layer of the hyper basis neural network which reconstructs dividing hypersurfaces between classes of arbitrary complex form. The characteristic property of the system is combined learning: matrix probabilistic neural network is tuned by lazy learning predicated on neurons at data point. Hyper basis neural network is tuned by controlled learning, herein, both synaptic weights and parameters of receptive fields are specified: centers and orientation of hyperellipsoids of activation functions. Along with high classification accuracy, the system is characterized by high learning speed which allows processing data streams that are fed sequentially in online mode.