| 引用本文: | 李鸿翔,王晓丽,阳春华,熊伟.基于GAN–UNet的矿石图像分割方法[J].控制理论与应用,2021,38(9):1393~1398.[点击复制] |
| LI Hong-xiang,WANG Xiao-li,YANG Chun-hua,XIONG Wei.Ore image segmentation method based on GAN–UNet[J].Control Theory and Technology,2021,38(9):1393~1398.[点击复制] |
|
| 基于GAN–UNet的矿石图像分割方法 |
| Ore image segmentation method based on GAN–UNet |
| 摘要点击 3113 全文点击 856 投稿时间：2020-08-24 修订日期：2021-02-06 |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| DOI编号 10.7641/CTA.2021.00558 |
| 2021,38(9):1393-1398 |
| 中文关键词 生成对抗网络 深度学习 矿石图像分割 |
| 英文关键词 generate adversarial networks deep learning ore image segmentation |
| 基金项目 国家自然科学基金项目(61673401), 国家自然科学基金基础科学中心项目(61988101)资助. |
|
| 中文摘要 |
| 在选矿生产过程中, 磨机给矿粒度对磨矿分级效率影响重大, 是一个关键的控制参数. 由于矿石表面不规
则、棱线较多, 同时存在矿石间堆叠的问题, 给基于图像的矿石粒度检测带来极大困难. 本文提出一种基于GAN–
UNet的矿石图像分割方法, 针对矿石图像棱线易引起矿石边缘错误识别的问题, 采用生成对抗网络进行图像分割,
将U–Net作为图像分割生成器网络, 使用人工标记的矿石边缘图像作为真实图像, 随后构建判别器网络以判断图像
来源, 同时将判别器误差与生成器误差通过加权形式引入网络训练中, 直到判别器难以判断分割图像来源, 获得满
足条件的生成器. 对实际工业生产矿石图像的分割结果表明, 本方法与U–Net网络相比提升了网络对矿石边缘的识
别能力, 减小了图像分割误差, 对矿石区域的相对误差平均值降至8.20%. |
| 英文摘要 |
| In the beneficiation process, the particle size of the ore has significant influence on the efficiency of grinding
classification process, and it is a key parameter for control. Due to the irregular surface of the ore, many ridges, and the
problem of stacking between ores, it is difficult to obtain accurate ore areas in image-based ore size detection methods.
A GAN–UNet based ore image segmentation method is proposed in this study. Considering the problem that there are
many edges in the ore image, it is easy to cause wrong recognition of ore edge, generative adversarial net is used for
image segmentation, the U–Net is used as the image segmentation generator network, using images with artificially marked
ore edges as real images. Then a discriminator network is constructed to determine the source of the image. At the
same time, the discriminator error and generator error are introduced into the network training in a weighted form. A
generator that meets the conditions will be obtained until it is difficult to determine the source of the segmented image by
the discriminator. Comparing to the U–Net network, the segmentation results by using actual industrial ore image show that
this method improves the ability to recognize the ore edges and reduces the error of image segmentation, and the average
value of relative error for ore area is reduced to 8.20%. |
|
|
|
|
|