Improving Tire Pattern Recognition Using Parameter-Efficient Fine-Tuning Techniques

Abstract

Tire-tread classification plays a key role in forensic investigation and public safety. This work introduces a robust, efficient recognition system that integrates Discrete Wavelet Transform (DWT) with Weighted Local Gray-Level on Robust Local Binary Pattern (WLG-RLBP), followed by weighted image fusion, which is subsequently processed by YOLO-family detectors (v5, v8, and v9). For class-incremental learning without access to old-class images, we adopt a unified label space and Label Distillation with Recall-Guard (LD), in which a frozen teacher produces pseudo-labels for old classes on new images via two stages: a high-confidence pass and a recall-oriented quota with IoU de-duplication, enabling supervision without retaining legacy data. We also employ Parameter-Efficient Fine-Tuning (PEFT), which freezes the backbone and updating lightweight neck/head adapters, reducing training time and memory. Evaluation spans two datasets (a public tire-impression set and a researcher-curated community set). As a reference baseline, full training on the public set yields mAP up to 87.73% with YOLOv8. Under LD + PEFT, the student models maintain high accuracy (mAP up to 94.1% with YOLOv8) while substantially lower computational cost. Grad-CAM visualizations focus on tread structures rather than background, supporting the method’s forensic credibility.