Rapid object recognition is traditionally explained by the initial feedforward sweep through the ventral visual pathway, yet increasing evidence suggests that later-stage processing becomes important when visual input places greater processing demands on the visual system. Previous studies have primarily investigated these demands by manipulating image properties such as occlusion or visual clutter, limiting the ecological validity of the stimuli. 

The present study implements a model-driven approach to identify natural images that differ in their processing demands without relying on predefined image manipulations or hand-crafted image characteristics. Two image sets were selected from the Natural Scenes Dataset using representational similarity analysis (RSA) by maximizing and minimizing the representational geometries between two convolutional neural networks (CNNs) that differed in their effective processing depth. 

Participants viewed these images during a rapid serial visual presentation (RSVP) task with and without backward masking while electroencephalography (EEG) was recorded. Time-resolved pairwise decoding and cluster-based permutation testing were used to compare neural representations across conditions and to assess whether any observed differences could be explained by low-level image statistics. Together, these analyses evaluate whether model-optimized image selection predicts differences in neural processing.