Synthetic Data Generation Techniques for Enhancing Crack Detection in Railway Concrete Sleepers

Developing deep learning–based computer vision systems for the autonomous inspection of railway sleepers is a challenging task due to the difficulty in obtaining training images representative of the actual infrastructure. To address this issue, this work proposes two different synthetic data generation methodologies. The first method involves a neural network based on Stable Diffusion 1.5, a type of latent diffusion model (LDM) capable of generating photorealistic images through a process that sequentially applies denoising autoencoders. Fine-tuning of this model was performed using images of real railway tracks to generate pictures of damaged concrete sleepers. The second proposed method is a Crack Simulator based on image processing techniques. It modifies images of undamaged sleepers by introducing textures and morphologies of cracks, which are randomly generated using one-dimensional Perlin noise signals. Additionally, it incorporates semantic classification maps of track elements to determine the origin and propagation zones of the cracks, considering factors such as ballast stone occlusions. The main objective of this study is to evaluate the capabilities of these synthetic approaches and compare their effectiveness with respect to the use of generic open-source databases. To this end, synthetic image data sets were created, and several YOLO (you only look once) object detection models were trained. The performance of the detectors was evaluated, and a quantitative assessment of each synthetically generated image data set was conducted. Results showed that the Stable Diffusion model exhibits remarkable potential in improving the concrete damage detection performance of deep learning systems.