|
Title |
A Stable Diffusion-Based Approach for Enhancing Blind Super-Resolution Performance of Rebar and Reinforcement Images
|
|
Authors |
심승보(Seungbo Shim) ; 양엄지(Eomzi Yang) |
|
DOI |
https://doi.org/10.11112/jksmi.2025.29.5.29 |
|
Keywords |
암맹 초해상화; 안정적인 확산 기법; 철근 및 배근 영상; 합성 데이터; 딥러닝 Blind super-resolution; Stable Diffusion; Rebar images; Synthetic data; Deep learning |
|
Abstract |
In recent construction sites, images of rebar and reinforcement often suffer from various degradation factors, such as low resolution, motion blur, and compression loss, as well as random noise introduced by the surrounding environment. These factors lead to the loss of structural information in the images, thereby reducing the accuracy and reliability of artificial intelligence-based analysis. In particular, due to the nature of construction sites, it is difficult to secure sufficient high-quality reference images, resulting in a lack of training data necessary for super-resolution algorithm development.
To address this limitation, this study proposes a novel blind super-resolution method that utilizes the Stable Diffusion X4 model to generate sharp, high-quality synthetic images from low-resolution images, which are then used together with real rebar images as training data. The proposed method develops a super-resolution algorithm by simultaneously using low-resolution images and synthesized high-resolution images as inputs. Experiments were conducted by applying the method to three state-of-the-art super-resolution neural network models. As a result, experiments conducted by applying the proposed method to three state-of-the-art super-resolution neural network models demonstrated that, compared to the conventional training approach using only real low- and high-resolution image pairs without synthetic data, the proposed strategy achieved an average improvement of 0.45 dB in resolution metrics and 0.85% in structural similarity metrics, while more accurately restoring the shapes and fine structural details of rebar and reinforcement. This study demonstrates that even in construction site environments where image quality is degraded, synthetic data can be effectively utilized to improve super-resolution performance. Furthermore, it is expected to contribute to enhancing the reliability of rebar condition analysis and advancing automation technologies for image-based structural management in the future.
|