Visual Pattern Recognition Using Enhanced Visual Features and PSD-Based Learning Rule

Abstract

This paper proposes a feedforward visual pattern recognition model based on a spiking neural network (SNN). The proposed model mainly includes four functional layers: 1) feature extraction; 2) encoding; 3) learning; and 4) readout. A modified HMAX model is first presented to extract features from external stimuli. In order to reduce the computational cost, we simplify the S1 layer using a single Gabor filter window. To simulate biological vision’s sensitivity to the vertical direction, we strengthen the feature of filtered orientation in 90° by adding a sharpened replica of the filtered image in 90° before max pooling in C1 layer. Then the phase encoding approach is used to convert the extracted visual features into spike patterns. These spike patterns will be learned by precise-spike-driven-based learning rules in an SNN. Finally, experimental results on benchmark datasets including MNIST, Caltech 101, and optical characters demonstrate the efficiency and robustness in noisy environments of the proposed model.