Research And Implementation Of Shape Feature Extraction Algorithm For Image Recognition

Shape is an advanced visual feature in the image.Due to its intuitiveness and interpretability,it is often used to represent the structure and attitude of the object.The extraction of shape feature is the foundation of shape analysis and recognition,and the extracted shape features will directly affect the subsequent algorithms on computer vision tasks.In this thesis,the development of the algorithms on extracting shape features is first reviewed.And the existing methods of shape descripors are suffering from the following drawbacks: 1)they are unable to capture the global information of the object;2)they can not well distinguish the objects with similar shape,which belong to different kinds of objects;3)there is no method that could precisely differ the edges from the shapes of the target object;4)usually the feature dimension are large and it consumes a lot of computation cost.In order to solve the above problems,this dissertation investigated both the global and local features of the object contour shape,and defined two new shape descriptors.Experiments results show that our methods achieve better performance.The main contribution of this thesis are summarized as follows:1.A low-dimensional multi-scale local feature is proposed.Based on the local features such as TAR and MTCD,a new contour descriptor is constructed,called the Margin Feature.It can well capture the edge information of the object and can measure the convexity/concavity of each point at different scales.It effectively solves the problem of distinguishing the objects with similar shape,which belong to different kinds of object.In addition,the feature dimension of Margin Feature is small.2.A novel global feature is proposed.At present,most classical shape descriptors focus on local features,and ignore the capture of global features of the contour.In this thesis,a contour global feature is proposed,call the Shape Feature.It can capture rich global information about the object's shape and can efficiently reflect the geometric properties of the shape,and its extraction process is simple and efficient.3.The fused features that combines the merits of popular shape descriptors and our proposed shape descriptor is implemented.At present,the mainstream shape features can capture the local information of the object shape but not enough,and they all ignore the global information.Therefore,combining the Margin Feature and Shape Feature that are proposed in this thesis together with the popular shape features can enable the feature descriptors to comprehensively describe the characteristics of the object's shape.In this thesis,the contour shape-based features extraction method is studied in depth.The disadvantage of the existing shape feature extraction methods are analyzed.Two shape descriptors are proposed and experimental verification have been carried out on specific application problems.Finally,the experimental results show that the proposed shape features outperforms the state-of-the-art features in the application of plant leaf recognition and QR code detection.