Monitoring corrosion in industrial components is crucial for ensuring operational safety, maintaining structural integrity, and preventing substantial economic losses. However, conventional automatic corrosion detection methods that rely on traditional image processing frequently encounter difficulties in capturing subtle surface alterations and gradual material deterioration. More advanced deep learning approaches are able to improve the detection rate. However, these methods generally rely on supervised learning, which requires the labor-intensive annotation of large, and often specialized datasets. This limits the method's practical application in optical structural health monitoring. This concern is especially relevant in the context of industrial heritage settings, which require high-individual training data.

This work addresses these challenges by leveraging deep learning techniques that combine latent space analysis with the CLIP (Contrastive Language–Image Pre-training) model. This approach quantifies corrosion progression in a scalable and annotation-efficient manner. By mapping images into a high-dimensional semantic latent space, our method identifies characteristic visual corrosion features, detects subtle changes over time, and enables robust, similarity-based tracking of material degradation.

The proposed framework allows visualization of corrosion development, extraction of representative corrosion feature vectors, and derivation of a relative corrosion metric by integrating image and text encoders. This approach facilitates the quantitative evaluation of corrosion and supports interpretability by connecting semantic latent representations to visually meaningful features.

The results demonstrate the potential of semantic latent space analysis to transform image date of corrosion into actionable insights for industrial asset management. This method provides a practical path toward continuous, automated corrosion monitoring, reducing dependency on manual inspection and enabling efficient, data-driven predictive maintenance strategies.