Stereoscopic Visual Attention Model Based on Bioinspiration

A stereoscopic visual attention model predicts the regions that people focus on most when viewing stereoscopic images, holding significant application value in the fields of robot vision, autopilot, virtual reality, augmented reality, and so on. Existing stereoscopic visual attention models are constructed based on neural networks to fit simply eye fixation datasets, and the globally effective receptive field (ERF) has not been carefully studied in these works. This article proposes a brain-like global stereoscopic attention model (BGSAM) from a bioinspiration perspective by utilizing the brain-inspired neural network CORnet-S as the backbone network. Following an exploration of ERF properties of dilated convolutions, we design an attention-atrous spatial pyramid pooling module to expand the ERF globally. Moreover, the BGSAM incorporates the disparity processing mechanism of the human brain based on a disparity energy model. Experiment results demonstrate that the BGSAM achieved a correlation coefficient and similarity of 0.86 and 0.742, respectively, on the SIS-3D dataset.