Research on the crack detection method of black coating based on machine vision and deep learning.

Abstract

In practical applications of the black high-radiation coating on the surface of porous materials, thermal stress can lead to the formation of micro-cracks on the surface, which may compromise the overall structural integrity and safety. This study proposes a machine vision sampling system to address the challenge of low-contrast imaging of small cracks in black coatings, affecting real-time detection accuracy. The system investigates the effects of various lighting methods on crack-background contrast. Additionally, it performs data augmentation and annotation on collected images to construct a dataset for black coating crack target detection. A BCC-YOLO crack detection algorithm is introduced, which builds upon the YOLOv10s model by incorporating an ADown module to replace traditional Conv and SCDown down-sampling modules, reducing both the number of model parameters and computational complexity while enhancing the model's feature extraction capability for small cracks. Furthermore, an iEMA attention mechanism module is integrated into the small-target detection layer, which combines the iRMB module with the EMA attention mechanism. This fusion maintains effective attention while reducing the number of parameters. The UIoU loss function replaces CIoU to accelerate convergence and improve training stability. Experimental results demonstrate that, compared to YOLOv10s, BCC-YOLO achieves improvements of 9.7%, 11.2%, 10.8%, and 9.8% in precision (P), recall (R), mAP₅₀, and mAP_50:95, respectively, on the self-built black coating crack dataset. Moreover, it reduces FLOPs by 7.3%. These enhancements improve the accuracy of crack detection for low-contrast black coatings and decrease the computational complexity of the model, which holds significant implications for achieving high-precision automatic crack detection.