Enhancing marble image classification performance via super-resolution-assisted image improvement

Abstract

Marble is one of the most widely preferred natural stones in various sectors such as construction, decoration and art, due to its aesthetic structure, durability and wide range of colors. This widespread usage has elevated marble beyond merely being a decorative material, turning it into a raw material with significant economic value. However, maintaining this economic value requires more than just high-quality physical and chemical properties; it is also crucial that the product images provided by manufacturers are clear, informative, and visually appealing. In this context, applying super-resolution (SR) methods to enhance marble imagery represents an innovative step toward bridging the gap between visual quality and automated digital evaluation. Nevertheless, marble images used in production and marketing processes are often of insufficient quality due to factors such as low resolution, blurriness or inadequate lighting. Poor visual quality reduces competitiveness, particularly in digital sales and promotional activities. This study aims to reconstruct low quality marble images using Super Resolution methods to enhance image quality. Images of Elazg Cherry Marble were used, forming a dataset comprising 2.551 image patches derived from 370 labeled slabs previously utilized in quality classification tasks. This ensured sufficient data diversity for model training and evaluation. Various super-resolution models, including GAN-based (e.g., ESRGAN) and attention driven (e.g., RSMAN) architectures, were employed. After SR reconstruction, the success of the enhancement was evaluated using image quality assessment metrics, namely PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index). The results demonstrated that images subjected to super-resolution were of higher quality and exhibited significantly improved visual clarity compared to the original low-resolution images. Moreover, classification tests conducted using the enhanced images achieved an accuracy rate of 96,4%. This indicates a positive impact not only on image quality but also on model performance. Consequently, it is concluded that this approach enhances customer satisfaction and strengthens competitive advantage within the marble industry. Overall, the integration of SR techniques not only improves visual perception but also enhances the functional performance of automated marble classification systems, leading to higher competitiveness and customer trust within the digital marketplace. Although this study focused on Elazg Cherry Marble, the proposed SR classification framework is inherently material-agnostic. It can be transferred to other natural stones or textured materials (e.g., granite, travertine, wood, textiles) with minimal domain-specific retraining, indicating strong potential for broader industrial adaptation.