Generative adversarial network (gan) based data augmentation for enhancing dl models on fa ade defect identification

Current façade safety inspections are done manually by relying on visual cues and have proven to be labor-intensive and unsafe. Deep learning (DL) techniques, particularly, convolutional neural network-based approaches, demonstrated great success in automated defect detection on surfaces. Yet, these models need large amounts of images for high performance. Collecting and labeling a large number of façade defect images is expensive and time-consuming, inducing data scarcity and imbalanced dataset problems. Previous studies aimed to improve DL models' accuracy with various methods but by training them with imbalanced and relatively small façade defect datasets. This study introduces a data augmentation approach using a deep convolutional generative adversarial network (DCGAN) to generate synthetic images of façade defects, addressing these issues. We evaluated the model using Frechet inception distance (FID) and visual inspection. Our study provides a data augmentation method to generate a much-needed dataset for training automated façade defect classification models.