Edge-assistant visual objects decoding using sparse representation

Visual objects decoding with functional magnetic resonance imaging (fMRI) often merely depends on the brain activity, using pattern classification to decode information about visual stimuli from patterns of activity. However, the spatial resolution of fMRI is still limited to the > mm range. Limited by its spatial resolution, fMRI voxels lack high-spatial-frequency information of visual stimuli. The overcomplete sparse representation can effectively match the sparse coding strategy in the primary visual cortex of human. Thus, it provides a probable way to represent the high-spatial-frequency information. Aiming to improve classification performance of images which were visual stimuli presented to participants, this paper proposed an edge-assistant approach for visual stimuli decoding, using sparse representation to supplement the ensemble of early visual voxels with high-spatial-frequency information. We supplemented early visual voxels with the sparse representation coefficients of edge patches and significantly improved the classification performance in a four-category (car, face, building, and animal) object classification analysis, which has valuable reference for practical fMRI-based image retrieval system.