Due to the multi-layer sealed encapsulation structure of the insulation layer and outer metal protective layer in medium- and low-temperature pressure pipelines, the corrosion area is completely visually obscured. Simultaneously, signal attenuation occurs during the transmission of penetrant testing signals through multiple layers of heterogeneous media. This results in acquired corrosion images exhibiting low contrast and weakened edge gradients, leading to low segmentation accuracy for local pitting corrosion cores and circumferential corrosion boundaries. To address this, a segmentation algorithm for corrosion areas beneath the insulation layer of medium- and low-temperature pressure pipelines is proposed. Pulse eddy current technology penetrates the insulation layer to collect electromagnetic response signals. Spatial integration of induced voltages combined with pseudo-color mapping enables visualization of corrosion beneath the insulation, resolving the visual obstruction issue. An improved multi-scale Retinex algorithm employs logarithmic domain decoupling and multi-scale Gaussian weighting estimation to separate low-frequency background from high-frequency details in the visualized image. Adaptive linear stretching is introduced to expand the grayscale dynamic range, enhancing corrosion features weakened by signal attenuation. Based on enhanced corrosion features, construct a two-dimensional histogram correlating pixel grayscale with neighborhood average grayscale. Determine the optimal segmentation threshold by maximizing inter-class dispersion to precisely delineate boundaries between localized pitting corrosion and circumferential corrosion. Experimental results demonstrate: Under low-temperature conditions of -10 °C and a signal-to-noise ratio of 19 dB, the algorithm achieves a pixel accuracy of 0.957 and a mean intersection-over-union ratio of 0.956 for corrosion image segmentation beneath insulation layers. It clearly delineates the core of pitting corrosion and the boundaries of circumferential corrosion, effectively resolving the challenge of precise corrosion region extraction under dual constraints of visual occlusion and signal attenuation. This provides a highly robust technical solution for detecting concealed corrosion in in-service pressure pipelines.
D'AguiarS C M, MottaR D S, AfonsoS M B.An investigation on the collapse response of subsea pipelines with interacting corrosion defects[J].Engineering Structures, 2024, 321: 1.1-1.14.
XieRuo-han, FanZheng, LiZhen, et al. Research on pipeline corrosion prediction coupled with artificial intelligence[J]. Applied Chemical Industry, 2025, 54(11): 2894-2898.
TianCong-wen, LiBo, LanWen-fei, et al. Semantic segmentation method for urban underground pipeline defects based on FCC-Deeplabv3+[J]. Journal of South-central Minzu University (Natural Science Edition), 2025, 44(1): 107-117.
MaoXin-yue, WangHui-feng, ZhouJia-le, et al. Semantic segmentation of phased array ultrasound images based on improved U-Net3+[J]. Journal of East China University (Science and Technology), 2025, 51(2): 242-249.
ZhangChi, ZhaoQiang, ZhangDa-zhen, et al. Research on collaborative segmentation method of large and small models for drainage pipe defect detection images[J]. Acta Scientiae Circumstantiae, 2025, 45(11): 28-35.
CaiQin, LiJi-cheng, QiZheng-wu, et al. Study on wall thickness measurement of pressure pipe with coated layer based on pulse eddy current (transient electromagnetic method) technology[J]. Hot Working Technology, 2023, 52(2): 21-26.
ChenTao, HeQian, ChengLyu, et al. Study on the hybrid detection method of pressure pipeline defects with EMAT/PECT[J]. Chinese Journal of Sensors and Actuators, 2023, 36(6): 860-867.
LiuWei-hua, XueYan-song, YiChen, et al. Low-light image enhancement algorithm combining multi-scale deep learning networks and retinex theory[J]. Journal of Signal Processing, 2023, 39(1): 105-115.
ZouYi-wen, LuXiao, WangLu-cai, et al. Low-light image enhancement based on Retinex theory and diffusion model[J]. Journal of Electronic Measurement and Instrumentation, 2025, 39(9): 182-191.
WuJia-yun, WuLing-zhi, HuXiao-fei. Multi threshold image segmentation based on general learning equalization optimizer[J]. Chinese Journal of Sensors and Actuators, 2024, 37(3): 463-468.