ESRGAN-LS: super-resolution reconstruction of multi-scale defects in aerospace thermal barrier coatings via attention enhancement and dynamic feature optimization

Performance breakthroughs and safety assurance in flight equipment remain pivotal drivers of technological advancement, with every smooth flight relying on the stable output of core propulsion systems and the reliable operation of critical components. As the key protective system for hot-end components in aeroengines, thermal barrier coatings (TBCs) are prone to multi-scale defects under extremely high-temperature and high-pressure operating conditions. These defects severely compromise the reliability and lifespan of hot-end components, thereby impacting flight safety. Addressing the limitations of existing X-ray and infrared thermal imaging technologies in resolution, as well as the high cost and low efficiency of scanning electron microscopy, this study proposes ESRGAN-LS-an enhanced generative adversarial network-based super-resolution reconstruction method for multi-scale defects in TBCs. The model incorporates the CL-SE attention mechanism to introduce the improved dense residual block RRDB-LS, enhancing local feature perception and reconstruction capabilities for multi-scale defects. While preserving the core framework, an additional RRDB layer is added, and the final eight RRDB layers are replaced with the enhanced dense residual block RRDB-LS module. This achieves shallow-layer preservation of fundamental texture reconstruction while comprehensively capturing multiscale features in images or data through the enhanced effect of local defect detection in the latter eight layers, enabling dynamic optimization of feature channels. Using the DIV2K dataset as the training set, experimental results demonstrate that, compared to the ESRGAN generator model, the ESRGAN-LS generator model achieves a 0.3% and 0.12% improvement in PSNR and SSIM metrics, respectively, on the Urban dataset. And on the Manga109 dataset, PSNR and SSIM improved by 0.35% and 0.1%, respectively. This significantly enhances visual discernibility and texture detail in defect regions, effectively preserving critical information at defect locations and markedly improving denoising performance after reconstruction. Two defect characterization experiments were conducted: defect ROI feature extraction and 3D visualization of defect features. Focusing on defect edge contours, these experiments significantly advance the intelligent recognition of multi-scale defects. The results validate the method's effectiveness in quantitative defect characterization and intelligent detection, concluding that the improved model designed in this study demonstrates superior reconstruction capabilities for texture details and edge information in multi-scale defects of aeroengine thermal barrier coatings. This research provides efficient technical support for ensuring engine safety and extending operational lifespan, holding significant implications for achieving high-precision visual inspection solutions in the high-end equipment manufacturing industry.