针对复杂背景下的目标识别,提出了一种同时检测目标位置和区分目标类别的识别方法.该方法首先从图像中提取丰富的兴趣点,通过图像之间的对极几何约束,过滤出精确匹配的图像兴趣点;然后在兴趣点特征空间用局部类别一致k均值聚类方法生成特征码本;最后,对于给定的测试兴趣点集,通过投票得到表示目标类别的局部特征,采用最大化目标类别和目标位置的联合概率得到前景目标.在Caltech-101数据库和实际场景图像上的实验表明,该方法的识别精度大约提高了8%.

Aiming at the object recognition in complex background,a recognition method,which simultaneously detects object location and classifies object types,is proposed.In this method,first,a large number of interest points are selected from the image and are filtered to obtain exactly-fitted outliers via the epipolar geometric constraint between images.Then,feature codebooks in the interest point space are formed via the local type consistent k-means clustering.Finally,for the given test interest points,the local features are voted to represent the object type,and the foreground object is obtained by maximizing the joint probability of object type and location.Experimental results on standard dataset Caltech-101 and real scene images demonstrate that the proposed method improves the recognition accuracy by about 8%.