Image augmentation for nondestructive testing in engineering structures based on denoising diffusion probabilistic model

doi:10.1016/j.jobe.2024.109299

Image augmentation for nondestructive testing in engineering structures based on denoising diffusion probabilistic model

2024-07-01

1

PRE

AI

DOI：10.1016/j.jobe.2024.109299

Original

Original request for help

Save

Abstract

En 中文

The performance of data-driven deep learning models relies on the quantity and quality of training data available. The scarcity of data in structural damage detection having been an important restriction for its rapid development. An image data augmentation method based on the denoising diffusion probabilistic model (DDPM) is proposed to focusing on structural damage detection task in this study. This approach employs a Markov chain sequence to progressively introduce random noise into the images, yielding isotropic Gaussian noise. By learning the reverse diffusion process, the original data is reconstructed from the Gaussian noise. Two sets of structural damage image data with varying complexities were utilized. Comparative studies were conducted against two widely used generative adversarial networks (GANs), including Wasserstein GAN-Gradient Penalty and Self-Attention GAN. the quality of generated images was evaluated using three metrics: inception score (IS), Frechet inception distance (FID), and kernel inception distance (KID). The experiments showed that the DDPM exhibits slightly better performance than GANs for low-complexity concrete crack image data. Specifically, it generates smooth images, whereas GAN-generated images contain more noise. The IS, FID and KID under DDPM are 2.45, 42.41 and 0.0326 respectively, while the best evaluation metrics under GANs are 2.38, 49.15 and 0.0377 respectively. For the high-complexity building damage image, the DDPM remains capable of generating high-quality images, while GANs can only generate images that are cluttered and severely distorted. It is shown that DDPM has significant advantages over GANs for augmenting high-complexity structural damage image data. Therefore, the proposed method can be applied to structural damage image data with different complexity, which provides a new method of high-quality data augmentation to address the problem of lack of training data in the field of deep learning-based intelligent damage detection.

Keywords:

Deep learning

Denoising diffusion probabilistic model

Image generation

Data augmentation

Intelligent damage detection

Image augmentation for nondestructive testing in engineering structures based on denoising diffusion probabilistic model

Abstract

Journal

Organization