Deep Learning-Based Semantic Segmentation of Thermal Defects Using AResU-Net and REAL-ESRGAN for the Infrared Image Resolution Enhancement

The detection of temperature anomalies in infrared (IR) images is one of the primary challenges of IR imaging methods for condition monitoring and fault diagnosis in industrial infrastructure. The accuracy of the segmentation methods highly depends on the resolution and overall quality of the captured IR images. Sometimes, high-resolution imagery is not available to cover the desired region of interest (ROI) in the captured scenes. Furthermore, in robotics-based condition monitoring procedures, the captured IR images might be blurred due to object motion, camera shaking, and defocusing. In this context, our paper studied the application of an emerging data preprocessing technique called REAL-ESRGAN for enhancing the overall resolution of captured IR images by reconstructing lost information. The reconstructed data sets were used to improve the overall performance of the Attention Residual (ARes) U-Net architecture in pixel-level segmentation of thermal defects.