基于WCDPM模型的细粒度物体识别

Fine-grained Recognition Based on WCDPM Model

针对可变形部件模型(deformable parts model,DPM)同等对待各部件,无法体现不同部件对识别过程的贡献度差异的不足,提出一种权重系数可变形模型(weighted coefficient deformable parts model,WCDPM),对DPM中的各部件赋予权重,强调区分度较高的部件在识别过程的作用,弱化区分度低的部件对识别的影响,提高细粒度识别精度.同时给出了模型的训练过程和权重系数的学习方法.在Airplan OID和Oxford-IIIT Pet两个数据集上进行实验,验证了该方法的有效性.

Since i t treats the parts equally,while the deformable parts model （ DPM） cannot highlight distinctive parts that are helpful to distinguishing subtle categories. To cope with the problem mentioned above,a weighted coefficient deformable parts model （WCDPM） was proposed to highlight distinctive parts and decrease the influence of non-distinctive parts,which leaded to improving performance in terms of fine-grained recognition accuracy. The detailed processes of model training and coefficient learning were also presented. Experimental results of Airplan OID and Oxford-IIIT Pet data sets demonstrate the effectiveness of the proposed method.