Orientation-Frequency Fields and Periodic Prior Reranking for Robust Cross-Modal Tire Indentation Retrieval

Cross-modal tire indentation retrieval refers to the task of retrieving corresponding tire surface images from a gallery using indentation images as queries. This task is inherently challenging due to factors such as low contrast, directional misalignment, local deformation, and other environmental disturbances. To address these challenges, this paper proposes a novel framework that integrates representation learning with geometric priors to enhance matching robustness and accuracy. During training, Orientation-Frequency Fields (OFF) are employed to guide the backbone network in capturing discriminative features related to tread orientation and pitch. During inference, a Periodic Prior Reranking (PPR) module is introduced, which performs normalization, period estimation, cyclic crosscorrelation for phase alignment, and period-wise tokenization for structural consistency verification, with the results fused through similarity scoring. To improve practical applicability, a phase-correlation-based coarse filtering module is incorporated to reduce the candidate search space. In addition, feature and periodic caches are adopted to eliminate redundant computations during retrieval. Finally, mask fallback, de-duplication, and tail completion strategies are integrated to further enhance robustness under challenging real-world conditions. On the test dataset, the proposed method achieves a Hit@10 accuracy of 75.81, outperforming the baseline model by a margin of 5.59. These results demonstrate that the proposed framework provides an effective solution for cross-modal tire indentation retrieval.