Attention choice on quotient space composition for automatic target image segmentation

The authors present a novel method for image segmentation using neocognitron neuron model with diffusion and concentration properties controlled by quotient structure. The contributions of this paper are two fold: (1) In order to remark upon the relationship between the components of an image corresponding to the patches, which give the typed association of objects, the notion of quotient space composition and projection is introduced. This provides the advantage that the new patches for image segmentation may be created by merging the patches from the above level of the quotient space. (2) The hierarchical structure of neocognitron for the attention choice is adjusted with the variance of scenes by the process of the diffusion and concentration. The validity of our approach is demonstrated by an example in image analysis.