Defect Detection in Tire X-Ray Images: Conventional Methods Meet Deep Structures

Andrei Cozma, Landon Harris, Hairong Qi, Ping Ji, Wenpeng Guo, Song YuanarXiv 2024 · DOI · OpenReview

TL;DR

This paper proposes a hybrid approach for automated tire defect detection in X-ray images, combining traditional feature engineering (LBP, GLCM, Fourier, Wavelet) with a Random Forest classifier. This method achieves competitive results against deep learning models like YOLOv8, demonstrating the continued value of well-engineered features, especially for complex industrial imaging tasks.

What Problem Does This Paper Address?

Ensuring tire integrity is critical for vehicle safety, and X-ray imaging is a key inspection method. However, manual inspection of these images is prone to human error, fatigue, and inconsistency, leading to missed defects or false alarms. The inherent challenges in tire X-ray images, such as anisotropic textures, varying defect sizes and shapes, low contrast, and complex tread patterns that can mimic defects, make automated detection difficult. Existing deep learning approaches often struggle with high-resolution images and data imbalance, prompting the need for robust and efficient automated solutions.

Methods

The core of this work lies in a hybrid approach that leverages both traditional feature engineering and modern machine learning.

  • Feature Engineering: The authors meticulously select and combine several traditional feature extraction techniques:
  • Local Binary Patterns (LBP): Captures local texture by comparing pixels with their neighbors, useful for identifying textural anomalies.
  • Gray-Level Co-occurrence Matrix (GLCM): Analyzes spatial relationships between pixels to understand texture uniformity and variations.
  • Wavelet Features: Decomposes images into multi-scale frequency components, ideal for detecting defects at different scales.
  • Fourier Features: Transforms images into the frequency domain to reveal underlying patterns and anomalies. Statistical features (mean, median, variance, etc.) are extracted from each of these.

  • Classification: A Random Forest (RF) classifier is used for the traditional method, chosen for its balance of accuracy and efficiency. Hyperparameters are optimized using grid search and 5-fold cross-validation.

  • Hybrid Approach with Deep Learning: To explore synergy, the engineered features are integrated into a YOLOv8 deep learning framework. This is achieved by augmenting the original grayscale image with additional channels representing processed versions of the engineered features (e.g., background removed images, wavelet reconstructions).
  • Probability Map Ensemble: For the RF-based approach, a probability map ensemble is constructed. Each image is divided into overlapping windows, and the RF classifier predicts the probability of a defect for each window. These probabilities are then aggregated and processed to create a final heatmap highlighting potential defect regions.

Datasets

The study utilizes a dataset collected from Sailun Group Co. Ltd., consisting of 1054 tire X-ray scans. These grayscale images are annotated with bounding boxes for two primary defect types: blisters and wire defects. The dataset presents significant challenges due to the wide variation in image sizes (width: 1536-3328px, height: 1625-14600px) and the small relative size of defects compared to the overall image dimensions.

Key Insights and Ideas

  • Traditional Features Still Matter: The research strongly suggests that traditional feature engineering, when carefully selected and combined, can yield highly competitive results, even against state-of-the-art deep learning models like YOLOv8. This challenges the notion that deep learning is always the superior solution for complex industrial imaging.
  • Synergy is Key: Combining engineered features with deep learning frameworks (as demonstrated by augmenting YOLOv8 inputs) can enhance performance, suggesting a promising direction for hybrid models.
  • Feature Importance: The study highlights that not all traditional features are equally beneficial. LBP alone performed poorly, and its inclusion in combinations sometimes diminished performance. Conversely, Wavelet features consistently improved performance, especially when LBP was excluded.
  • Dataset Preparation is Crucial: The effectiveness of the RF model is significantly influenced by dataset preparation techniques, particularly window size, step size, and the threshold for determining ground truth overlap (IOMA). Larger windows and a higher overlap threshold generally led to better results.

Key Results

  • Random Forest Performance: The best-performing traditional approach, combining GLCM, Fourier, and Wavelet features with an RF classifier, achieved high precision (0.922%), recall (0.918%), and F1-score (0.918%) in per-window classification.
  • Comparison with YOLOv8: While YOLOv8 models showed some improvement with augmented features, the optimized RF approach with engineered features often matched or surpassed their performance in terms of precision and recall, particularly for the "Blister" class. For instance, the Random Forest (G,W) method achieved 0.759% precision and 0.766% recall for blisters, outperforming the baseline YOLOv8.
  • Optimal Configuration: For the RF model, a window size of 384 pixels, a step size of 32 pixels, and an IOMA threshold of 0.3 yielded the highest performance metrics (Precision: 0.973%, Recall: 0.980%, F1: 0.976%).

Limitations and Future Work

  • Limitations:
    • The study acknowledges the challenge of data imbalance, where defect-free samples far outnumber defective ones, which can bias models.
  • The anisotropic nature of tire X-ray images poses a challenge for standard CNN architectures.
  • The authors note that the YOLOv8 models used in their comparison were not necessarily the absolute best-performing configurations, and further tuning might yield different results.
  • Future Work:
    • The authors aim to expand the framework's applicability and further refine its performance.
  • They advocate for continued exploration of hybrid approaches that combine traditional feature engineering with advanced machine learning to tackle complex industrial automation tasks.
  • Future work will focus on further optimizing dataset generation parameters and exploring more advanced feature combinations.

Summary updated

Abstract

This paper introduces a robust approach for automated defect detection in tire X-ray images by harnessing traditional feature extraction methods such as Local Binary Pattern (LBP) and Gray Level Co-Occurrence Matrix (GLCM) features, as well as Fourier and Wavelet-based features, complemented by advanced machine learning techniques. Recognizing the challenges inherent in the complex patterns and textures of tire X-ray images, the study emphasizes the significance of feature engineering to enhance the performance of defect detection systems. By meticulously integrating combinations of these features with a Random Forest (RF) classifier and comparing them against advanced models like YOLOv8, the research not only benchmarks the performance of traditional features in defect detection but also explores the synergy between classical and modern approaches. The experimental results demonstrate that these traditional features, when fine-tuned and combined with machine learning models, can significantly improve the accuracy and reliability of tire defect detection, aiming to set a new standard in automated quality assurance in tire manufacturing.
BibTeX
@misc{cozma2024defect,
  author = {Andrei Cozma and Landon Harris and Hairong Qi and Ping Ji and Wenpeng Guo and Song Yuan},
  title = {{Defect Detection in Tire X-Ray Images: Conventional Methods Meet Deep Structures}},
  year = {2024},
  eprint = {2402.18527},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2402.18527},
}
Figure