Infrared and Laser Engineering, Volume. 54, Issue 7, 20250035(2025)

Visual inspection of surface crack morphology of low-contrast coatings based on improved U-net

Shaohua CHEN1,2, Shida ZHANG1,2, Jiaojiao REN1,2, Jian GU1,2, and Lijuan LI1,2
Author Affiliations
  • 1Key Laboratory of Photoelectric Measurement and Optical Information Transmission Technology of Ministry of Education, School of Opto-Electronic Engineering, Changchun University of Science and Technology, Changchun 130022, China
  • 2Zhongshan Institute of Changchun University of Science and Technology, Zhongshan 528400, China
  • show less

    Objective During actual use of the black high-emissivity coating on the surface of porous materials, micro-cracks on the surface of micron scale are generated due to thermal stress. In order to facilitate the subsequent development of the evolution and expansion laws of crack defects under thermal stress, low-contrast coating surface crack morphology research on visual inspection technology. A detection method that combines optical optimization and deep learning is proposed. By designing a light source to stimulate a monocular vision system, first optimizing the illumination method and incident angle parameters from the perspective of system design to enhance the local contrast of collected crack images. It proposes an algorithm to adapt to low-contrast image crack contrast enhancement. In turn, an improved U-Net network is built to improve the ability to extract low-contrast crack features by embedding attention modules, deep hyperparametric convolution and activation functions. Experimental results show that the local contrast of the acquired images is the highest when the incident light is 30° in the high illumination mode. After preprocessing, the image contrast is increased from 10.507 to 42.662, which effectively reduces the influence of background noise on crack information when the image is low contrast, and can better highlight the morphological characteristics of cracks. The Dice coefficient, SSIM index and accuracy Acc of the improved network reached 0.862, 0.892, and 0.901 respectively in terms of crack segmentation performance indicators. The detection rate of cracks with widths greater than 9.6 μm reached more than 90%, and the crack shape and direction were clearly recognizable.Methods In order to identify the shape and direction of microcracks on low-contrast images and define the minimum detection width, this study built a monocular vision acquisition platform to collect high-quality images starting from the lighting method and light incidence angle (Fig.1), so as to improve the contrast between microcracks and the substrate background, preprocess the image to reduce the influence of the background on crack information, and improve the contrast between cracks and background (Fig.4). Through the improved U-Net network model (Fig.11), an attention mechanism is added at the junction of down-sampling and up-sampling jumps to avoid the model being affected by image noise, improve the extraction ability of key features, and use deep hyperparametric convolution to increase the number of convolution kernels., more features can be extracted, thereby improving the model's representation ability and segmentation accuracy to complete the detection of cracks on the coating surface, and realizing crack segmentation.Results and Discussions Based on the design of the vision system in this research, high-quality images were obtained. The contrast of the original image was 10.507, and after pre-processing, the contrast was increased to 42.662, which was improved by 4 times, make the difference between the crack information and the background area in the image more obvious. Through this algorithm Att-Do-U-net combines the attention mechanism with the deep hyperparametric convolution structure, it ultimately performs the best among various indicators. The highest values were reached in terms of Dice coefficient of 0.862, SSIM index of 0.892 and accuracy of 0.901. In addition, in terms of segmentation results, the segmented crack information has a more complete vein structure, and also has a good segmentation effect on small branches of cracks (Fig.16). The lines are continuous and smooth on the segmentation results, which is better than the results of broken lines, breakpoints, etc. in other results. The cracks that cannot be detected are discussed (Fig.18). Between 4.8 and 9.6 μm, the crack detection rate is less than 25%, while between 9.6 and 14.4 μm, the crack detection rate is as high as more than 90%. For crack widths above 14.4 μm, the detection rate reaches 100%, so the minimum width of detectable cracks is defined as 9.6 μm.Conclusions In this paper, a light source excited monocular vision system is constructed. By optimizing lighting methods and image preprocessing algorithms, combined with improving U-Net network, accurate detection of low-contrast micro-cracks is achieved. The pretreatment method in this paper can reduce the interference of noise and obtain complete crack information. The segmentation effect of low-contrast crack images based on the improved U-net network is better than that of the original U-shaped network. The crack segmentation is more complete. The boundary of crack morphology characteristics on the result is smoother. The average SSIM reaches 0.892, and the Dice coefficient reaches 0.862. A crack width of 9.6 μm can be recognized in terms of crack width definition. The current method has limitations for crack detection less than 10 μm. In the future, super-resolution reconstruction technology can be introduced to recover the crack skeleton from low-resolution crack images to achieve lower-width detection.

    Keywords
    Tools

    Get Citation

    Copy Citation Text

    Shaohua CHEN, Shida ZHANG, Jiaojiao REN, Jian GU, Lijuan LI. Visual inspection of surface crack morphology of low-contrast coatings based on improved U-net[J]. Infrared and Laser Engineering, 2025, 54(7): 20250035

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Infrared

    Received: Jan. 13, 2025

    Accepted: --

    Published Online: Aug. 29, 2025

    The Author Email:

    DOI:10.3788/IRLA20250035

    Topics