This study aims to detect pathologies in constructions made of sillar, a volcanic material of great historical and cultural value, commonly used in residential and heritage buildings, in the city of Arequipa, Peru. Due to the uniqueness of sillar and the particular characteristics of its pathologies, such as moisture and surface wear, a non-invasive methodology using digital images is proposed, oriented toward the analysis of heritage constructions, with the objective of developing a method that does not alter or modify the heritage or damage the structure, considering that in invasive studies, sample collection may affect the integrity of the material. The proposed strategy combines computer vision techniques, including clustering methods for preliminary segmentation, with the use of deep neural networks for anomaly and deterioration detection. Furthermore, a validation scheme is introduced that integrates standard segmentation metrics with intersection analysis relative to pathology maps, allowing computational analysis to align more closely with the criteria employed in architectural conservation. The results demonstrate good performance in moisture detection, although with lower accuracy in identifying other types of deterioration, highlighting both the feasibility and the challenges of applying deep learning to sillar diagnostics and laying the groundwork for the development of digital tools that support the documentation and preservation of architectural heritage.

The Historic Centre of Arequipa, a UNESCO World Heritage Site, contains buildings constructed mainly of sillar, a white volcanic ignimbrite, whose ornamented façades are time-consuming to document. This study evaluates whether image-to-image translation models can convert façade photographs into two-dimensional façade representations. A dataset of 578 heritage façades was collected through field photography and Google Street View, preprocessed through obstacle removal, homography-based rectification, and stitching, and paired with reference representations produced with generative tools and corrected by architects. Three models (Pix2Pix, Pix2PixHD, and Contrastive Unpaired Translation, CUT) were trained and evaluated on an independent test set of 66 façades using the mean absolute error (MAE), structural similarity index measure (SSIM), and peak signal-to-noise ratio (PSNR). Pix2Pix obtained the best scores (MAE = 7.24, SSIM = 0.768, PSNR = 22.90 dB), followed closely by Pix2PixHD, whereas CUT performed worse (MAE = 20.80, SSIM = 0.418, PSNR = 16.04 dB). An exploratory assessment of three buildings showed that 50–58% of the architectural components were successfully generated. The models reproduced the overall façade layout but not complex ornamentation. However, the limited availability of heritage-building data restricts the evaluation of model generalization.