期刊文献+

一种结合图像分割掩膜边缘优化的B-PointRend网络方法 被引量:1

B-PointRend: mask edge optimization method combined with image segmentation
下载PDF
导出
摘要 在实例分割的过程中,Mask R-CNN在CNN特征之上添加全卷积网络(FCN)来生成图像掩膜。其生成的掩膜是由28×28预测掩膜放大为检测框尺寸得到的,因而对于物体边缘不敏感,PointRend通过对实例边缘的"难点"进一步处理得到更精确的实例分割结果,但由于"难点"为两个类别的边界,所以并不容易确定类别。本文提出了一种改进的B-PointRend方法,该方法把图像的像素尺度信息加入到预测中,在实例分割流程之后增添一个边缘修正步骤,采用像素级别的图像边缘检测和处理方式,在分割结果尺度上对检测框掩膜的边缘进行修正。实验结果表明,改进方法得到的物体的掩膜边缘更光滑且连续,更精确地覆盖了真实的物体边缘,在实例分割的准确率等指标上均有提升。 In instance segmentation,Mask R-CNN adds a fully convolutional network(FCN)to predict the mask of each detected instance.The generated mask is obtained by upsampling a 28×28 instance mask to the size of detection frames,so it is not sensitive to the edge of objects.PointRend processes the“uncertain regions”of the instance edge to get more accurate instance segmentation results.However,due to“uncertain regions”is the boundary between two categories,it is not easy to determine the category.The upgraded B-PointRend method add the pixel scale information of an image to the prediction.An edge correction step is added after the instance segmentation process.The edge of the detection frame masks is corrected on the scale of the segmentation result using pixel-level image edge detection and processing.The experimental results show that the edges of object masks are smooth and continuous,and covers real object edges better.The accuracy of instance segmentation is improved.
作者 雷晓春 李云灏 梁止潆 江泽涛 LEI Xiaochun;LI Yunhao;LIANG Zhiying;JIANG Zetao(School of Computer Science and Information Security,Guilin University of Electronic Technology,Guilin 541004,China;Guangxi Key Laboratory of Image and Graphic Intelligent Processing,Guilin University of Electronic Technology,Guilin 541004,China)
出处 《中国体视学与图像分析》 2021年第3期261-268,共8页 Chinese Journal of Stereology and Image Analysis
基金 国家自然科学基金(No.61876049,61762066) 四川省区域创新合作项目(2021YFQ0002)
关键词 实例分割 B-PointRend 深度学习 掩膜 instance segmentation B-PointRend deep learning mask
  • 相关文献

同被引文献8

引证文献1

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
使用帮助 返回顶部