A Biologically Inspired Hierarchical Model Of Shape Coding For Object Recognition

The destination of computer vision is to make the computer recognize objects like humans.Applying the biological visual information processing mechanisms to object recognition is currently an active research focus.Shape features,which are the significant feature for describing objects,are hierarchically processed from V1,V2,V4 to IT visual cortex along the ventral pathway.Because of the complexity of shape features' processing mechanisms in cortex,there is not any perfect theoretical system yet.Traditional computer vision studies focus on V1 to extract bar features,which ignores the importance of high-level cortex in representing object shape.Applying the processing mechanisms of high-level cortex to extracting shape features is considerable and valuable.Considering that the bar features extracted from V1 are insufficient for representing object shape,we concentrate on the shape processing mechanisms in V1,V2 and V4 cortex,and present a new hierarchical model for shape extraction to extract angle and curvature features.First,Gabor is adopted to extract bar features.Then,3D-DOG inhibiting noise is combined with Gabor filters,which are different with 90 degrees,are introduced to extract angle features in V2.Finally,the computation of curvature field is proposed to describe deformation of object shape and extract histogram of curvature and gradient orientation,followed by a feature fusion.The hierarchical representation based on the bar,angle and curvature features help to enhance the key points of object shape.Moreover,the fusion features also improve the inadequacy of using only curvature or gradient feature for object description.Experiments on MNIST handwritten digits and 21 remote sensing images,compared with classical models inspired by biological vision and traditional computer vision models demonstrate that the proposed model can achieve desirable classification accuracy.At the same time,orientation expanding,scale expanding,hierarchical representation and features fusion are verified to be effective.