Bembenek Michał, Mandziy Teodor, Ivasenko Iryna, Berehulyak Olena, Vorobel Roman, Slobodyan Zvenomyra, Ropyak Liubomyr
Department of Manufacturing Systems, Faculty of Mechanical Engineering and Robotics, AGH University of Science and Technology, 30-059 Kraków, Poland.
Department of the Theory of Wave Processes and Optical Systems of Diagnostics, Karpenko Physico-Mechanical Institute of the NAS of Ukraine, 5 Naukova St., 79060 Lviv, Ukraine.
Sensors (Basel). 2022 Oct 7;22(19):7600. doi: 10.3390/s22197600.
This paper describes the combined detection of coating and rust damages on painted metal structures through the multiclass image segmentation technique. Our prior works were focused solely on the localization of rust damages and rust segmentation under different ambient conditions (different lighting conditions, presence of shadows, low background/object color contrast). This paper method proposes three types of damages: coating crack, coating flaking, and rust damage. Background, paint flaking, and rust damage are objects that can be separated in RGB color-space alone. For their preliminary classification SVM is used. As for paint cracks, color features are insufficient for separating it from other defect types as they overlap with the other three classes in RGB color space. For preliminary paint crack segmentation we use the valley detection approach, which analyses the shape of defects. A multiclass level-set approach with a developed penalty term is used as a framework for the advanced final damage segmentation stage. Model training and accuracy assessment are fulfilled on the created dataset, which contains input images of corresponding defects with respective ground truth data provided by the expert. A quantitative analysis of the accuracy of the proposed approach is provided. The efficiency of the approach is demonstrated on authentic images of coated surfaces.
本文介绍了通过多类图像分割技术对涂漆金属结构上的涂层和锈蚀损伤进行联合检测。我们之前的工作仅专注于在不同环境条件（不同光照条件、阴影存在、低背景/物体颜色对比度）下锈蚀损伤的定位和锈蚀分割。本文方法提出了三种损伤类型：涂层裂纹、涂层剥落和锈蚀损伤。背景、油漆剥落和锈蚀损伤是仅在RGB颜色空间中就可以分离的对象。对于它们的初步分类，使用支持向量机。至于油漆裂纹，颜色特征不足以将其与其他缺陷类型区分开来，因为它们在RGB颜色空间中与其他三类重叠。对于初步的油漆裂纹分割，我们使用谷值检测方法，该方法分析缺陷的形状。一种带有改进惩罚项的多类水平集方法被用作高级最终损伤分割阶段的框架。在创建的数据集上进行模型训练和准确性评估，该数据集包含相应缺陷的输入图像以及专家提供的各自的地面真值数据。对所提出方法的准确性进行了定量分析。该方法的效率在涂层表面的真实图像上得到了证明。
