Recurrent convolutional model based on gated spiking neural P system for stereo matching networks

The rapid development of deep learning techniques has introduced extensive research improvements to various aspects in the processing pipeline of the stereo matching problem. Due to the high requirements of 3D convolution on computing resources and the domain sensitivity of 2D convolution, some stereo matching networks have begun to shift from full convolutional structures to recurrent structures, using the hidden state update mechanism of recurrent units to achieve global consistency of disparity-related information. In this paper, a new recurrent convolutional model is constructed based on a two-dimensional spiking neural computational system, and three types of recurrent units are designed by setting different parameters. The newly designed recurrent units are applied to a recent recurrent stereo matching network for better disparity propagation. Starting from the definition of two-dimensional gated spiking neural P systems, the spiking mechanism of a single neuron is extended to multi-neurons arranged in a two-dimensional array which are locally topologically connected. Its state update mechanism is parameterized in a form that can be back-propagated, thus realizing a new type of recurrent convolutional model. The proposed model can be embedded into existing recurrent stereo matching networks. Experimental results demonstrate that it can effectively reduce the computational load of the baseline method and achieve comparable accuracy to existing state-of-the-art methods.