Hierarchical Representation Of Visual Scene

Human intelligence has an incomparable advantage over artificial intelligence systems in dealing with complex and varied external environments as it has the capacity of fast learning and flexible migration.The processing efficiency of intelligent system depends on its internal representation(Marr,1982;Neisser,1967).Good representations are an important reason why human intelligence has this advantage.Most of the recent visual perception research focused on psychological representations such as "feature"and "object"(Treisman&Gelade,1980;Kahneman,Treisman&Gibbs,1992).However,the reality is much more complicated than the "features" and "objects".The vision system is able to recognize multiple objects in the scene,understands their complex relationships,and extracts the semantics of the visual scene(Zhu&Mumford,2007).How to effectively represent complex and varied visual scenes in human cognitive systems is an important theoretical issue in the field of cognitive psychology.Since the hierarchical representation with visual grammar can construct a multi-level relationship by recursively to represent the whole scene and different scene information can be expressed in this form,I propose that the scene information is represented as a visual grammar tree.This paper intends to carry out systematic research about the above assumptions.This research adopts the psychophysical and computational modeling techniques,and systematically investigates the hierarchical structure representation of the scene region segmentation information with the depth effect of the nodes in the hierarchical structure as the index.The entire study consists of two parts.Study 1 mainly explored whether the visual system can construct hierarchical representation of multiple regions with complex spatial relationships in the scene.In this part,the hierarchical tree with the visual grammars which define the relations between the segmented regions was used to generate scene segmentation image.I compared the memory performance of different depth nodes in hierarchical tree to test whether the hierarchical representation was formed or not.In Study 2,a hierarchical model for the scene region segmentation with the horizontal and vertical cutting rules was built and tested.The model was set up with the estimated human prior and tested by comparing the performance of human and model in the tasks.The study has the following main conclusions:(I)For the region sepgmentation information in the scene,the changes of different depth nodes in the hierarchical structure produce different behavioral performances,which cannot be explained by the differences in the type of color change and size of color change area.The above results indicate that there is a hierarchical representation in the processing of scene information.(2)For the scene region information with different structure types and different segmentation grammars,the performance differences still exist between different depth nodes.It shows that,hierarchical representation is universal for the scene region information.(3)A cognitive model for the region segmentation process based on Bayesian inference is constructed.The model implements the process of parsing a scene region segmentation into a hierarchical tree and uses the tree to complete the follow-up computation.It achieves human-level performance in cognitive tasks.This study is a useful attempt to develop a psychological theory that describes the computational process and an exemplary case for applying psychology research to artificial intelligence algorithm design.