【摘要】 本文算法基于目标检测模型Mask RCNN进行改进,改进后的模型能分割和标记出不同的猪个体并且能精确分割出猪头区域。首先选用Resnet50作为模型的特征提取网络,然后考虑到群养猪图像分割任务的特殊性,在区域建议网络(Region Proposal Networks, RPN)中引入感兴趣区域(Region of Interest, ROI)的非局部特征向量,最后为进一步提高分割掩模边缘精度,在ROI输出的掩模分支中提出使用sobel检测滤波器预测目标边缘,并在损失函数中加入边缘损失。实验选取1000张图像作为训练样本,200张图像作为测试样本,结果表明,该算法模型在测试样本上对猪体和猪头分割的平均召回率达到0.851和0.845,相对Mask RCNN模型分别提高了6.4%和7.2%,并且在训练速度上相对Mask RCNN提高了18%。提取分割出的猪个体可进一步进行身份识别研究,精确分割出的猪头区域可用于饮水、吃食等行为识别。更多还原

【Abstract】 The algorithm based on target detection model Mask RCNN is improved. The improved model can segment and mark different pig individuals and accurately segment the pig head area. Firstly, Resnet50 is selected as the feature extraction network of the model, and then the non-local features of the Region of Interest(ROI) are introduced in the Region Proposal Networks(RPN) considering the particularity of the image segmentation task of the group pigs. Vector, and finally to further improve the edge precision of the segmentation mask, a sobel detection filter is used to predict the target edge in the mask branch of the ROI output, and edge loss is added to the loss function. Experiments selected 1000 images as training samples and 200 images as test samples. The results showed that the average recall rate of pig model and pig head segmentation on the test sample reached 0.851 and 0.845, which was 6.4% higher than the Mask RCNN model. And 7.2%, and the training speed is increased by 18% compared to Mask RCNN. The extracted pig individual can be further subjected to identification research, and the accurately segmented pig head area can be used for behavior recognition such as drinking water and eating.更多还原