利用多种特征和Hopfield神经网络的有遮挡的目标识别

Occluded Objects Recognition Using Multiple Features and Hopfield Neural Network

该文提出了一种新的识别有遮挡目标的方法 ,即将目标模型和含有目标的遮挡图象的轮廓在某一尺度上张角的极值、极值点 (亦称显著点 )之间距离、相对位置等信息集成在一起 ,作为描述目标模型 (遮挡图象 )的一组特征 ,且这组特征在平移、旋转和均匀尺度变换下保持不变 .其轮廓上点 p处的张角可用余弦定理很方便地求出 ,而张角的极值点则对应于轮廓急剧变化的地方 .同时将特征匹配定义为模型特征与遮挡图象特征之间的对应 ,若这种对应被映射到 Hopfield神经网络上 ,则该网络即可用于完成全局特征匹配 .该文提出的方法已在 P 个人计算机上用Matlab5 .2编程实现 ,并给出了应用实例 .实验结果表明 ,该方法能有效地从含有遮挡目标的景物图象中识别出目标 。

In this paper, we propose a new approach to recognize occluded objects. The information of the magnitude of the local extreme of the open angles in the contour of a model(occluded image) at a scale, the information of the distance and relative location between the two adjacent dominant points are suitably integrated as a set of features for describing a model(occluded image), the features are invariant under rotation, uniform scaling, and translation of the curve. The magnitude of opened angle at a point p i in the contour can be easily calculated by the law of cosines, and its local extreme correspond to the sharper changes of the contour of the mode(scene). The feature matching is to define the correspondence between the model features and the scene features. Each correspondence between a model feature and a scene feature constitutes a “feature correspondence pair”, they are mapped onto the Hopfield neural network that is used to perform global feature matching. The proposed approach has been implemented on PⅡ personal computer in Matlab5 2 programming language and examples are presented. The experimental results show that our proposed method can efficiently recognize an object from an image of occluded objects, and be implemented easily.