Cross-Modal Tire Impression Comparison via Knowledge Graphs and Adaptive Weight Learning

The effective matching of indentations with tire tread mark remains challenging with the advancement of artificial intelligence methods and paradigms. The delay in addressing this problem arises from multiple factors, including the design of the tread pattern, wear state, inflation pressure, and load conditions. Furthermore, the discrepancy between the conditions under which the dataset is collected and the conditions in which query images are obtained-where the latter are typically compromised due to collisions and other factors-exacerbates the issue. This paper presents a novel approach that leverages knowledge graphs and adaptive learning to improve performance in a crossmodal context. We introduce cross-domain recommendation and contrastive learning methodologies to extract features both within and across domains. A domain-specific knowledge graph is constructed to encode pattern types, indentation shapes, and physical compatibility as comparable scores. At the sample level, we integrate “feature quality” and “knowledge confidence” into an adaptive weight learning framework. This approach enables a stronger emphasis on visual similarity when matching clear images, while prioritizing knowledge scoring for fuzzy images. We conduct extensive experiments, achieving an overall accuracy of 85.4%, significantly surpassing the 63.7% accuracy obtained by traditional feature matching. Additionally, we perform scenariobased experiments to assess the method’s applicability in realworld settings.