高级检索+

基于实例分割和光流计算的死兔识别模型研究

Dead Rabbit Recognition Model Based on Instance Segmentation and Optical Flow Computing

  • 摘要: 为实现自动化识别死兔,提高养殖管理效率,以笼养生长兔为研究对象,以基于优化Mask RCNN的实例分割网络和基于LiteFlowNet的光流计算网络为研究方法,构建了一种多目标背景下基于视频关键帧的死兔识别模型。该模型的实例分割网络以ResNet 50残差网络为主干,结合PointRend算法实现目标轮廓边缘的精确提取。视频关键帧同时输入实例分割网络和光流计算网络,获取肉兔掩膜的光流信息和掩膜边界框中心点坐标。利用光流阈值去除活跃肉兔掩膜,通过核密度估计算法获取剩余中心点坐标的密度分布,通过密度分布阈值实现死兔的判别。实验结果表明,肉兔图像分割网络的分类准确率为96.1%,像素分割精确度为95.7%,死兔识别模型的识别准确率为90%。本文提出的死兔识别模型为兔舍死兔识别和筛选工作提供了技术支撑。

     

    Abstract: Screening and isolating dead rabbits is one of the important work of meat rabbit farms, which is helpful to build a rabbit breeding safety system. In order to identify dead rabbits automatically and improve the efficiency of breeding management, cage-rearing breeding rabbits was taken as the research object, a dead rabbit recognition model was proposed which was based on the modified Mask RCNN and LiteFlowNet. The instance segmentation part of the model used ResNet 50 residual network as the backbone, used PointRend algorithm as the network head to extract the instance contour accurately. The key frames of the rabbit videos were sent to rabbit instance segmentation network and optical flow calculation network at the same time to obtain the optical flow of the meat rabbit mask and the center point coordinates of the instance boundary boxes. The masks of the active rabbits were removed by the threshold of the optical flow, and then the density distribution of the remaining center point coordinates was obtained by kernel density estimation algorithm, and the dead rabbits were distinguished by density distribution threshold. The experiment results showed that the classification accuracy of the rabbit segmentation network was 96.1%, the pixel segmentation accuracy of the rabbit segmentation network was 95.7%, and the recognition accuracy of the dead rabbit recognition model was 90%. This study provided technical support for dead rabbit recognizing and isolating in rabbit farms.

     

/

返回文章
返回