Abstract: Local descriptors， such as SIFT （scale invariant feature transform）， encode the local appearance information of the interest points as image features with rotation， scale and viewpoint invariance. This method is widely used in many fields including object categorization， object recognition and image matching， etc. However， the biggest problem for most local descriptors is that they only capture the local information but ignore the geometric configuration， which is useful in describing the object. Structural context， a new local descriptor which encodes the geometric configurations between the interested points， is introduced. The experimental results prove that it is more powerful in describing the object category.

Key words: local descriptor, interest point, scale invariant feature transform （SIFT）, shape context, structural context

摘要： 局部描述符（如SIFT）方法能够将图像中关键点的局部表观信息作为图像的特征，具有旋转不变性、尺度变换不变性、仿射不变性等性质，被广泛应用于物体分类、物体识别、图像匹配等领域。但是，它存在一个重要缺陷：只能描述物体的局部特征，忽略了整个物体的构造，而这在表示物体时是非常重要的。设计了一个新的“结构上下文”局部描述符，通过当前关键点和其他关键点间的空间拓扑结构关系描述各个关键点的特征。实验证明这种描述符在描述相同物体种类时特别有效。

关键词: 局部描述符, 关键点, 尺度不变特征变换, 形状上下文, 结构上下文