<p>Accurate non-invasive detection of surface and subsurface defects is vital for cultural-heritage preservation. Yet conventional approaches are constrained by scarce data, environmental interference, and harsh conditions. Solar-loading thermography (SLT), a non-contact infrared method, reveals flaws in historic buildings and artworks. However, SLT’s dependence on sunlight and other factors often yields noise, hindering defect interpretation. To address this, SLT is combined with an enhanced YOLOv7-ECA model to improve defect detection in aged canvas paintings. The proposed methodology, validated on a laboratory canvas replica of “The Crucifixion of Sansepolcro,” includes attention enhancement to emphasize defects and suppress textures, noise-reduction, and COMSOL-based synthetic-data augmentation to expand datasets and improve robustness. Experiments show higher accuracy for deep-seated defects, including splitting and biological attack, even in challenging environments. Through this study, an automated pathway for non-contact, non-destructive testing in cultural-heritage conservation is established, directly addressing noise interference, data scarcity, and complex environmental adaptability during defect detection.</p>