LIGHTWEIGHT VISION TRANSFORMERS FOR DEFECT DETECTION IN LOW-QUALITY INFRASTRUCTURE IMAGES

Abstract

Accurate defect detection in infrastructure images is essential for timely maintenance, safety assessment, and cost-effective asset management. However, real-world inspection images are often affected by low resolution, poor illumination, motion blur, compression artifacts, occlusion, and complex background noise, which reduce the reliability of conventional computer vision models. This paper presents a lightweight vision transformer-based framework for improving defect detection in low-quality infrastructure images. The proposed approach focuses on identifying visible surface defects such as cracks, corrosion, spalling, leakage marks, and material degradation while maintaining computational efficiency for practical deployment on resource-constrained devices. Unlike heavy transformer models that require high memory and processing capacity, the lightweight design uses compact attention mechanisms, efficient feature extraction, and image-quality-aware learning to enhance defect representation under degraded visual conditions. Experimental findings indicate that the proposed model achieves improved detection accuracy, stronger generalization, and better robustness compared with baseline convolutional and transformer-based methods. The results further show that lightweight vision transformers can preserve fine defect details while reducing inference time and model complexity. This makes the approach suitable for field inspection, mobile-based monitoring, drone imagery, and automated infrastructure maintenance systems. Overall, the study demonstrates that efficient transformer architectures can provide a practical and scalable solution for reliable defect detection in challenging low-quality infrastructure images.