结合掩膜和可变形部件模型的扣件定位算法

Railway Fastener Location Algorithm Combined Mask and Deformable Part Model

针对原始可变形部件模型在定位铁路扣件时误定位率过高的问题,提出一种结合掩膜和可变形部件模型的扣件定位算法。在训练步骤中,对训练出的扣件可变形部件模型作掩膜操作,去除模型中的部分扣件背景区域。统计去除过后模型部件得分的分布,从而确定各个部件的权重。对于一张新的待定位图像,采用训练完成的部件权重计算模型检测得分,根据得分高低确定扣件位置。实验结果表明,与目前主流扣件定位方法相比,原始可变形部件模型定位效果较好,提出的算法相比于原始可变形部件模型误定位率减小了52%。另外,在强阳光和阴天时,模型同样能够准确定位扣件。

In order to solve the problem that the original deformable part model has high false locating rate when locating the railway fastener, this paper proposes a railway fastener location algorithm combined mask and deformable part model. First, in the process of training, the trained fastener deformation part model is masked to remove part of the fastener background area in the model. Then, the distribution of masked model part score is counted to determine the weight of each part. Finally, for a new image which is not located, the weight of the trained part is used to calculate the model detection score. Experimental results show that the original deformable part model has better locating effect than current mainstream fastener location methods and the proposed algorithm reduces the false locating rate by 52% compared to the original deformable part model. In addition, the proposed algorithm can also accurately locate the fastener on sunny days and cloudy days.