基于Mask R-CNN的马匹四肢别征提取方法研究与应用

Research and application of horse limbs feature extraction method based on Mask R-CNN

针对马品种登记中对马匹护照自动标记马匹外貌特征的需求,提出一种基于Mask R-CNN的马匹四肢别征的提取方法,对马匹四肢上别征进行精确检测和分割,并在马匹外貌特征图示中进行自动标记。利用新疆马业协会提供的马匹图像、网络爬虫和实地拍摄的图像,构建马匹四肢别征数据集,在Mask R-CNN框架上对单个模型与马匹模型、四肢模型和别征模型进行对照实验,4个模型的AP值分别为83.7%、91.5%、90.1%和89.2%。发现单个模型进行三层分割会使像素点存在歧义性,无法对马匹外貌特征进行自动标记。实验结果表明,使用3个模型联级进行检测与分割,效果显著,实现了马匹四肢别征的自动标记。

Aiming at the requirement of automatic marking horse appearance features in horse breed registration,a method of extracting horse limbs features based on Mask R-CNN is proposed.The goal is to accurately detect and segment the horse limbs features,and automatically mark them in the horse appearance feature diagram.Based on the photo of horses,web crawlers and on⁃site photos provided by Xinjiang Horse Industry Association,the data set of four limbs characteristics of horses was constructed.The control experiments were carried out on the framework of Mask R-CNN between single model,horse model,limbs model and the characteristic model.The AP values of the four models were 83.7%,91.5%,90.1%and 89.2%respectively.It is found that the three⁃layer segmentation of a single model will make the pixels ambiguity and can not automatically mark the appearance features of horses.The experimental results show that the detection and segmentation effects of the three models are remarkable,and the automatic marking of horse limb features is realized.