Research on an Improved Car Tire Defect Detection Algorithm based on YOLOv5s

Currently, most companies rely on X-ray imaging technology to acquire X-ray images of car tires for detecting defects. The detection process involves manual visual inspection or traditional machine learning methods, which are both labor-intensive, costly, and unreliable, resulting in high false detection rates. In this study, we propose an improved car tire defect detection method based on YOLOv5s. To address the performance degradation caused by imbalanced sample categories, we introduce a confidence loss function to reduce the network’s focus on easily classifiable objects. Additionally, to tackle the challenge of detecting tire defects with large scale variations, we incorporate an adaptive feature fusion module to enhance the network’s representational power and its ability to handle multi-scale features. As a result, the detection performance of the network is significantly improved. According to the test data, the proposed method achieves an overall detection accuracy of $96.4 \%$ for classifying and localizing tire defect components. The accuracy rates for detecting specific defects such as open defects, cord defects, impurities, and edge misalignment are $\mathbf{9 7. 2 \%}, \mathbf{9 0 \%}$, $\mathbf{9 9. 3 \%}$, and $\mathbf{9 9. 3 \%}$ respectively.