| 中文题名： | 基于scribble的弱监督图像语义分割 |
| 姓名： | |
| 保密级别： | 公开 |
| 论文语种： | 中文 |
| 学科代码： | 070101 |
| 学科专业： | |
| 学生类型： | 学士 |
| 学位： | 理学学士 |
| 学位年度： | 2021 |
| 学校： | 北京师范大学 |
| 校区： | |
| 学院： | |
| 第一导师姓名： | |
| 第一导师单位： | |
| 提交日期： | 2021-05-25 |
| 答辩日期： | 2021-05-15 |
| 中文关键词： | 图像语义分割 ; scribbles标签 ; 图像模型 ; 弱监督 ; 全卷积神经网络 |
| 中文摘要： |
在图像语义分割中，强监督需获取逐像素点标签，耗费巨大精力，本文主要介绍一种基于 scribbles 标注的图像分割方法，旨在解决难以获取逐像素点标签这一问题。通过构建超像素点，依据空间限制、外观、语义内容，将信息从 scribbles 传播到无标记像素。同时，学习了一个完全卷积的神经网络，它由传播得到的标签监督，进而完成语义分割任务。用超像素点外观及空间相似性作为一项指标，神经网络训练结果作为另一项，构造联合损失函数，并用交替训练的方法求解模型。最终通过在 Pascal VOC 数据集上的实验结果表明，该方法在节约了大量成本的情况下具有良好的准确率，与其他几种标签相比拥有更好效果。
﹀
|
| 外文摘要： |
The problem of image segmentation is a hot issue in the field of computer vision, but the acquisition of pixel-by-pixel labels consumes a lot of energy. This article mainly introduces an image segmentation problem based on scribbles annotation, which aims to solve the problem of difficulty in obtaining pixel-by-pixel labels. By constructing superpixels, label is disseminated from the known to the unknown. At the same time, a FCN network is learned, which is supervised by the propagated labels to complete the semantic segmentation task. The appearance and spatial similarity of superpixels is used as an indicator, and the training result of neural network is used as another, a joint loss function is constructed, and the model is solved by an alternate training method. Finally, the experimental results on the Pascal VOC data set show that this method has a good accuracy rate while saving a lot of costs, and has better results than other types of tags.
﹀
|
| 参考文献总数： | 27 |
| 插图总数： | 0 |
| 插表总数： | 0 |
| 馆藏号： | 本070101/21102 |
| 开放日期： | 2022-05-25 |
谷歌
