Lightweight Internal Damage Segmentation Using Thermography with and Without Attention-Based Generative Adversarial Network

A lightweight internal damage segmentation network (IDSNet) (Ali and Cha, Autom Constr 141:104412, 2022) segments internal damages in concrete using active thermography. This study further investigates the performance of IDSNet for segmentation of subsurface damage. The IDSNet consists of an intensive module and a superficial module. The intensive module focuses on contextual features and learning complex correlations, and the superficial module learns spatial features in the input. The lightweight IDSNet has only 0.085 million parameters and processes a thermal image of 640 × 480 × 3 with 74 frames per second. The deep learning network requires extensive data for training; therefore, attention generative adversarial network known as AGAN (Ali and Cha, Autom Constr 141:104412, 2022), was used to generate artificial data for training IDSNet. The IDSNet was trained with and without AGAN data. The IDSNet achieved a 0.767 average intersection over union (aIoU) without using AGAN data and 0.891 aIoU with using AGAN data.