一种基于子类别感知的可变形部件模型对象检测算法

Object Detection Algorithm Based on Subcategory-Aware Deformable Part Models

为解决可变形部件模型中基于长宽比的启发式子类别划分方法难以处理复杂的类内变化问题,提出一种基于形状特征的子类别划分方法,并将其应用到可变形部件模型的训练过程中。对于样本实例,首先基于Canny算子提取对应的边缘信息。提取多尺度力矩特征构建训练样本的形状特征并将其作为特征向量,采用模糊C均值聚类算法进行子类别初始化;然后针对每一个子类别训练一个组件模型形成混合组件模型,优化模型的表现能力。PASCAL VOC 2007数据集的实验结果表明提出的算法超出初始的可变形部件模型的检测性能,平均精度提高了3.5%。

To solve the problem that the heuristic subcategory split method based on the aspect-ratio of deformable part models couldn＇ t handle complex intra-class variations well, the paper proposes a subcategory split algorithm based on the shape teature and applies it to the training process of de- formable part models. Edge intormation is first extracted from sample instances based on Canny op- erator. Then the multi-scale Torque teature is used to construct the shape teature according to the edge information and the teature vector of that is applied for subcategory initialization based on Fuzzy C-mean clustering algorithm. A component model is trained for each subcategory to form the mixture component model to optimize the performance of the model. The experimental results on Pascal VOC 2007 dataset show that the algorithm in the paper outperforms the performance of original DPM, achieving an improvement of 3.5% MAP.