Shape Recognition Method Based On Feature Learning

Shape recognition is a basic problem in computer vision,and is widely used in many fields,such as object recognition,image registration.However in practical applications,there are various deformations such as projective and elastic.Sketch is a kind of shape,it is an abstract result of the corresponding object which is processed by human brain.Every person has its own understanding of object,visual of angle and habit of painting,so there are more complex deformations in sketch.Accordingly,how to extract features which are invariable before and after deformations has become burning issues in shape recognition fields.Basic shape descriptor can hardly represent the invariable features,so how to obtain invariable features based on learning method has become main trend.For recognizing shapes which undergoes projective deformations,most of methods regard relations between sample points as shape descriptors.Because of the projective deformation,it is difficult to compute the correspondences of sampling point.In order to overcome the shortcomings of these descriptors,we propose a new curvature based shape coding and recognition method.First,we use projective invariant to describe of basic features of contour fragments and encode them,this operation can obtain middle-level feature of shape;moreover,we divide contour fragments into some groups based on their curvatures and extract features of different groups,then combine these features into shape features.Experiments results show that our method surpasses other methods when recognizing shapes which undergoes projective deformation.At present,most of sketch recognition methods ignore the temporality of sketch.So in this paper,we propose a temporality based sketch recognition method.First,we divide strokes of sketch into stroke groups,then extract features of stroke groups;in order to make better use of temporality,we feed features of stroke groups to gated recurrent unit,and use joint bayesian to fuse features of different time steps,at last we use k-nearest neighbor to classify sketch.Our method fully take into account the various characteristics of sketch.And experiments show that our method can get better performance than state-of-art methods.