Research On The Neural Representations Of Different Levels Of Visual Scene Properties Based On Functional Magnetic Resonance Imaging

Visual scenes contain complex information and different levels of properties.Although recent computer vision develops rapidly in scene recognition due to the rise of deep neural network,its dependency on large quantities of annotated samples and long training time limit its practical applications.However,human can understand scenes rapidly and accurately.Therefore,investigating the cognitive and neural mechanisms of human scene understanding can further provide references and inspirations to computer vision.This study used functional magnetic resonance imaging(fMRI)to investigate the role of mid-level scene properties,objects and low-level scene properties,spatial frequency in the human neural mechanisms of scene understanding.In the first experiment,we investigated the neural mechanisms when signature objects were masked in scene categorization.Significant changes were observed in the interactions between the object-selective region,lateral occipital complex(LOC)and the attention-related regions in the default mode network(DMN),and top-down modulation effect from the dorsolateral prefrontal cortex(DLPFC)to LOC and the extrastriate visual cortex,as well as the functional connectivity within a network of which the core hub locates in the left middle temporal gyrus(MTG).The results suggest that masking of signature objects significantly affected the object attention,cognitive demand,top-down modulation effect,and semantic judgment.In the second experiment,we investigated the role of low spatial frequency(LSF)and high spatial frequency(HSF)information in scene category representations in natural and indoor scenes under different conditions of luminance contrast.The results showed luminance contrast can significantly affect the role of spatial frequency in scene perception.In addition,we observed the natural scene recognition relied more on LSF information,while indoor scenes relied more on HSF information.The distinct spatial frequency preferences may be due to the different strategies of global and local perception in different scene categories.Based on the different levels of visual scene properties,this study is not only the first to directly reveal the interactions between the high-level cognitive regions and visual regions in the neural mechanisms of object-based scene categorization,but also indicates that low-level properties can carry crucial information in the understanding of mid-level spatial properties and further act on high-level semantic comprehension.These findings can help to improve the complete hierarchical structure in scene perception and processing of humans.