联合特征码本树和能量最小化的目标识别方法

Visual Object Recognition Method Based on Feature Vocabulary Tree and Energy Minimization

针对复杂背景下的目标识别,提出了一种基于特征码本树和能量最小化的概率框架,能同时检测目标位置和区分目标类别的识别方法。为了能加入特征间的空间关系,除了使用单特征码本树,还使用了双特征码本树,并建立一个能量函数来融合单特征码本树和双特征码本树的特征概率匹配结果。最后,通过在测试图像中寻找滑动窗口所在区域的类别能量最小化来确定目标的位置和所属类别。在UIUC和Caltech 101数据库上的实验表明,该方法相比于其他方法具有较高的识别精度。

Aiming at the object recognition in complex background,a object recognition method based on feature vocabulary tree and energy minimization,which simultaneously detects object location and classifies object types,was proposed.In order to join the spatial relationships between features,the vocabulary trees of single feature was used and pairwise feaure was concatenated.Then an energy function was proposed to combine the matched type probability results of these two vocabulary trees.At last the visual object location and object type are known by minimizing the energy function of the sliding window in test image.Experimental results on standard dataset UIUC and Caltech 101 demonstrate the proposed method has much higher recognition accuracy than other methods.