Using biologically inspired visual statistics to predict texture memorability

We might remember an image for its meaning, but purely visual features may also contribute. Although some image properties that increase visual stimulation (such as size and contrast) are known to enhance memory, the role of an image's local spatial structure remains unclear. We used a computational model of early vision to test whether the patterns of activity evoked by a texture predict how well human observers remember it. Orientation- and spatial-scale-selective receptive fields were combined with colour-opponency mechanisms to quantify population luminance and colour processing of a large set of texture images with minimal semantic content. We used the model to select stimuli that offered either high or low visual stimulation. One hundred observers rated how 'memorable' each image appeared (metamemory rating), then completed a memory test distinguishing previously seen images from novel ones (recognition). Textures evoking stronger activity were associated with higher metamemory ratings and enhanced recognition performance. However, while pattern and colour contrast statistics explained a large proportion of the variance in metamemory ratings (suggesting an influence of early visual processing), they explained less variance in recognition performance, indicating a smaller influence on recognition. These findings highlight which visual signals are likely to influence memory (and our perception of it) within a neurobiologically plausible feature space.