Adaptive large-kernel convolutional network with gaussian-modulated spatial calibration for tyre quality assessment

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Introduction Automated tyre defect classification is critical for ensuring vehicle safety and manufacturing quality in the automotive industry. Although deep learning has made significant strides in industrial visual inspection, existing convolutional networks struggle to simultaneously capture fine-grained local defect patterns and long-range contextual dependencies in tyre surface images. Methods To address this limitation, we propose UniConvNet, an adaptive large-kernel convolutional network that integrates the OverLoCK Context-Mixing Dynamic Convolution (ContMix) with a novel Gaussian-modulated spatial calibration mechanism. The proposed architecture progressively expands the receptive field through a dual-branch design: a multi-scale channel split branch extracts hierarchical local features, while an OverLoCK ContMix dynamic kernel branch adaptively models long-range dependencies. A Gaussian Distribution Modulation module enhances feature robustness against illumination variations, followed by Spatial Feature Calibration to suppress background clutter. Furthermore, a Progressive Large Kernel Regulation strategy 9 × 9 → 11 × 11 balances local detail preservation with global context aggregation. Multi-scale feature fusion aligns and compresses multi-stage outputs for the final classification head. Results Extensive experiments on the TyreNet dataset (1,698 images) and a public Roboflow tyre defect dataset (3,069 images) demonstrate that the proposed method achieves 96.45% classification accuracy on TyreNet with only 4.2 GFLOPs and 31.4 M parameters, outperforming ResNet-50, ConvNeXt-T, and OverLoCK-T while maintaining competitive efficiency. Discussion Cross-dataset validation confirms strong generalization, highlighting the practical applicability of the proposed approach for industrial deployment.