【摘要】 在马匹鉴定及马匹护照生成的过程中,需要对马匹面部的特征进行文字描述和描图标记。存在工作量大、效率低、准确性较低、人工成本较高等问题。为解决这些问题,提出了一种基于Mask R-CNN框架对马匹面部及马匹面部特征识别并进行实例分割的方法。首先通过网络爬虫和马场实地拍摄获取马脸正视图像,利用Labelme制作数据集标签,通过ResNet-50-FPN的主干网络从特征金字塔的不同级别提取RoI特征,其中的RolAlign算法,使用双线性内插的方法,提高了RoI定位精度,最后通过RoI单独应用的边界框识别和掩码预测,生成相应的马匹面部及马匹面部特征掩码,实现图像中马匹面部及马匹面部特征与背景的分割。此外,构建了一个具有分割标注信息的马脸数据集,用于训练相应模型。实验结果表明,该方法具有较好的马脸检测效果,并能在准确检测的同时实现像素级的马匹面部及马匹面部特征信息分割,可为马匹鉴定和护照生成提供一定支持。更多还原

【Abstract】 In the process of horse identification and horse passport generation, it is necessary to describe and mark the characteristics of the horse’s face. There are problems such as large workload, low efficiency, low accuracy and high labor cost. Therefore, we propose a method based on Mask R-CNN framework for horse face recognition and horse facial feature recognition and instance segmentation. Firstly, the face image of the horse is obtained through web crawlers and field shooting, dataset labels are made by Labelme, and RoI features are extracted from different levels of the feature pyramid through the backbone network of ResNet-50-FPN. Among them, the RollAlign algorithm uses bilinear internal interpolation method to improve the accuracy of RoI positioning. Finally, the corresponding mask of horse face and horse facial features is generated through the boundary box recognition and mask prediction separately applied by RoI to realize the segmentation of horse face and horse facial features from the background in the image. In addition, a horse face dataset with segmented labeling information is constructed for training the corresponding model. The experiment shows that the proposed method has better horse face detection effect and can achieve pixel-level horse face and horse facial feature information segmentation while accurately detecting, which can provide support for horse identification and passport generation.更多还原