Abstract
In recent years, research on decoding brain activity based on functional magnetic resonance imaging (fMRI) has made eye-catching achievements. However, constraint-free natural image reconstruction from brain activity remains a challenge, as specifying brain activity for all possible images is impractical. The problem was often simplified by using semantic prior information or just reconstructing simple images, including digitals and letters. Without semantic prior information, we present a novel method to reconstruct natural images from the fMRI signals of human visual cortex based on the computation model of convolutional neural network (CNN). First, we extracted the unit output of viewed natural images in each layer of a pre-trained CNN as CNN features. Second, we transformed image reconstruction from fMRI signals into the problem of CNN feature visualization by training a sparse linear regression to map from the fMRI patterns to CNN features. By iteratively optimization to find the matched image, whose CNN unit features become most similar to those predicted from the brain activity, we finally achieved the promising results for the challenging constraint-free natural image reconstruction. The semantic prior information of the stimuli was not used when training decoding model, and any category of images (not constraint by the training set) could be reconstructed theoretically. We found that the reconstructed images resembled the natural stimuli, especially in position and shape. The experimental results suggest that hierarchical visual features may be an effective tool to express the human visual processing.