A neural model of scene understanding: Multiple-scale spatial and feature-based attention in scene search, learning, and recognition

This dissertation develops neural models of how the brain learns to use multiple-scale visual information to efficiently search and recognize scenes and the objects within them.;The first project introduces ARTSCENE, a neural classifier based on principles of biological vision and categorization. Consistent with human psychophysical data, ARTSCENE embodies coarse-to-fine visual processes whereby spatial attention is deployed to multiple scales of information, from global gist to local textures, to learn and recognize scenic properties. Specifically, the model uses scene gist to generate a rapid hypothesis of scene identity, and then accumulates evidence from scenic textures to refine this hypothesis. The model shows how texture-fitting allocations of spatial attention, called attentional shrouds, can facilitate scene recognition, particularly when they include a border of adjacent textures. Tested on a benchmark photograph dataset, the ARTSCENE system classifies each testing image into one of four landscape scene categories (coast, forest, mountain and countryside) with up to 91.85% correct, outperforms alternative models in the literature that use biologically implausible computations, and outperforms component systems that use either gist or texture information alone.;The second project considers how salient objects and their spatial configuration in a scene constitute predictive contexts that facilitate rather than distract target search. The proposed model, ARTSCENE Search, implements associative learning of locations or identities between a target and its surrounding objects, and later uses such knowledge to predict target locations or identities, which are named spatial and object cueing, respectively. It follows the ARTSCENE framework in that global scene layout at the first glance rapidly forms a hypothesis of target locations, and sequential eye-scans to local objects incrementally refine the hypothesis by enhancing target-like objects in space. ARTSCENE Search simulates the interactive dynamics of spatial and object cueing in the cortical 'What' and 'Where' pathways starting from early visual areas through medial temporal lobe to prefrontal cortex. The model explains challenging psychophysical data of contextual cueing effects, and clarifies the functional roles whereby different brain areas, including the perirhinal and parahippocampal cortices, may coordinate scene perception, learning, and memory.