Mold infestation poses a persistent threat to the preservation of paintings on paper, yet current detection practices remain largely manual and difficult to scale. This study introduces the Mold Segmentation Network (MSN), a lightweight 15-layer convolutional neural network designed for automated, pixel-wise segmentation of mold defects in high-resolution scans. To support model development, a dedicated Mold Image Dataset (MID) was compiled from two ink-on-paper artworks, yielding 100 expertly annotated mold‑bearing images for training, validation and testing. MSN was benchmarked against three established architectures (FCN, U-Net, and SegNet). On the held‑out test set, MSN achieved higher sensitivity and Intersection over Union (IoU) scores than all comparison models while maintaining competitive overall accuracy, indicating superior recovery of subtle, diffuse mold regions. These findings demonstrate that a compact CNN tailored to mold morphology can serve as an effective, highly accessible triage tool for early biodeterioration detection and conservation monitoring in heritage collections.
Razak et al. (2026) studied this question.