Improvement of Damage Segmentation Based on Pixel-Level Data Balance Using VGG-Unet
Improvement of Damage Segmentation Based on Pixel-Level Data Balance Using VGG-Unet
复制标题
DOI：
10.3390/app11020518
复制
发表时间：
2021-01-01
影响因子：
2.7
通讯作者：
Suzuki, Yasuhiro
中科院分区：
文献类型：
--
作者：
Shi, Jiyuan;Dang, Ji;Suzuki, Yasuhiro
在这项研究中，收集了 200 张钢铁腐蚀图像和 500 张橡胶支座裂纹图像并手动标记以构建数据集。然后分别采用这两个数据集以两种方法训练VGG-Unet模型，旨在通过输入不同大小的数据集进行损伤分割。一种方法是挤压分割，将高分辨率的挤压图像直接输入 VGG-Unet 模型，而裁剪分割则使用尺寸为 224 x 224 的裁剪图像作为输入图像。由于数据集中损伤像素的比例不同，因此两个数据集产生的结果有较大差异。对于大尺寸损伤（如腐蚀）分割，裁剪分割具有较好的效果，而对于较小尺寸损伤（如裂纹）分割，结果相反。主要原因是数据集中有效数据的集中度存在差距。为了提高基于裁剪分割的裂纹分割能力，采用背景数据丢弃率（BDDR）来减少背景图像的数量，以在像素级上控制数据集中损坏像素的比例。数据集中损坏像素的比例可以通过不同的BDDR值来决定。通过测试，在BDDR为0.8的情况下，Cropping Segmentation的准确率相对较高。
In this research, 200 corrosion images of steel and 500 crack images of rubber bearing are collected and manually labeled to build the data set. Then the two data sets are respectively adopted to train VGG-Unet models in two methods, aiming to conduct Damage Segmentation by inputting different size of data set. One method is Squashing Segmentation to input squashed images from high resolution directly into VGG-Unet model while Cropping Segmentation uses cropped image with size 224 x 224 as input images. Because the proportion of damage pixels in the data set is different, the results produced by the two data sets are quite different. For large size damage (such as corrosion) segmentation, Cropping Segmentation has a better result while for minor damage (such as crack) segmentation, the result is opposite. The main reason is the gap in the concentration of valid data from the data set. To improve the capability of crack segmentation based on Cropping Segmentation, Background Data Drop Rate (BDDR) is adopted to reduce the quantity of background images to control the proportion of damage pixels from the data set in pixel-level. The ratio of damage pixels from the data set can be decided by different value of BDDR. By testing, the accuracy of Cropping Segmentation becomes relatively higher under BDDR being 0.8.