A Novel Pixel-Wise Defect Inspection Method Based on Stable Background Reconstruction.

Here are the English translations of the author information in Markdown format:

The authors of this paper include:

- **Chengkan Lv**: IEEE member, Bachelor of Science from Shandong University, Ph.D. Doctorate from the Institute of Automation, Chinese Academy of Sciences, research interests include neural networks, computer vision, and anomaly detection.
- **Shen Fei**: Master of Engineering, Research Fellow at the Strong Magnetic Field Science Center of the Chinese Academy of Sciences, supervisor of master's students, research interests include image feature extraction, pose measurement, and robotic dolphins.
- **Zhang Zhengtao**: Research Fellow at the Institute of Automation, Chinese Academy of Sciences, and the University of Chinese Academy of Sciences, research interests include feature extraction and defect detection.
- **Xu De**: Research Fellow at the Institute of Automation, Chinese Academy of Sciences, research interests include feature extraction, visual control, image processing, precision assembly, and humanoid robots, with multiple patents and awards for scientific and technological progress.
- **Yonghao He**: Bachelor of Software Engineering from Sichuan University, currently a Ph.D. candidate at the National Key Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, research interests include data mining, computer vision, and machine learning, especially the application of deep learning in multimedia retrieval.

Abstract

• A defect detection method for industrial product texture surfaces based on background reconstruction is proposed.
• The method consists of two modules:
  • An autoencoder integrated with Generative Adversarial Networks (GAN) to reconstruct defect-free texture backgrounds as references.
  • A detection network based on U-net for pixel-level analysis, comparing the differences between the original image and the reconstructed image.
• Experimental results demonstrate the effectiveness and generalizability of the method.

I. Introduction

• The importance of surface defect detection in industrial production.
• The limitations of traditional manual detection methods.
• Challenges of defect detection methods based on deep learning: lack of defect samples.
• The advantages of anomaly detection methods: no need for defect samples.
• Classification of anomaly detection methods: hyperplane construction and background reconstruction.
• Limitations of background reconstruction methods: unstable reconstruction and noise in normal regions.
• Contributions of this paper:
  • Designed a background reconstruction method based on AE and GAN, incorporating abnormal images to improve reconstruction stability.
  • Used U-net for defect detection based on the differences between reconstructed and original images.

II. Method

• A. Motivation and Overall Framework

  • Limitations of AE-based background reconstruction: unstable reconstruction and noise in normal regions.
  • Solutions proposed in this paper:
    • Introduce abnormal images to train the anomaly mapping method, improving reconstruction stability.
    • Use U-net for pixel-level difference analysis, suppressing noise in normal regions.
  • Overall framework illustration.

• B. Stable Background Reconstruction Based on AE and GAN

  • AE integrated with GAN to enhance the clarity of reconstructed images.
  • Introduce abnormal images to train the anomaly mapping method:
    • Use an auxiliary classifier C_A to distinguish reconstructed images.
    • Use auxiliary loss L_assist to optimize the reconstruction of abnormal images.
  • Network structure illustration.

• C. Difference Analysis Network Based on U-Net

  • Use U-Net for pixel-level difference analysis to segment defect regions.
  • Train with artificially generated defect images:
    • Generation algorithm: Randomly generate multiple line segments on defect-free images and replace them with abnormal images.
    • Use mask images as the ground truth (GT) for the segmentation network.
  • Network structure illustration.

III. Experiments and Analysis

• A. Datasets and Evaluation Metrics

  • Use the MAD dataset for experiments.
  • Evaluation metrics: AUC.

• B. Impact of Anomaly Image Mapping Method

  • Experimental results: The proposed method effectively removes defects, while AE + GAN cannot completely remove defects.
  • Limitations of the proposed method: Performance is limited for images with a large number of unique local structures.

• C. Impact of Different Anomaly Images

  • Experimental results: Combining multiple types of abnormal images for training can improve detection performance.

• D. Ablation Study

  • Experimental setup: Ordinary U-net, No anomalies, Proposed method.
  • Experimental results: The proposed method achieves the best performance on all texture images.

• E. Comparison with Other Methods

  • Compare with 8 commonly used anomaly detection methods:
    • AE, MSCDAE, AnoGAN, f-AnoGAN, CNN Feature Dictionary, Texture Inspection, Iterative Projection, FCDD.
  • Experimental results: The proposed method achieves the best performance on all texture images.

• F. Production Line Experiment

  • Apply the proposed method to defect detection in mobile phone cover glass.
  • Experimental results: The proposed method effectively removes defects and suppresses noise in normal regions.

IV. Conclusion

• This paper proposes a defect detection method based on background reconstruction, which does not require defect samples.
• The proposed method can effectively remove defects and suppress noise in normal regions.
• The proposed method achieves the best performance on a variety of texture images.

Q: What specific research methods were used in the paper?

• Background Reconstruction: Utilizing the combination of Autoencoders (AE) and Generative Adversarial Networks (GAN) to reconstruct the original images into defect-free reference images.
• Anomaly Mapping: Introducing additional anomaly images during the training process and training an auxiliary classifier to distinguish between reconstructed normal images and anomaly images, thereby guiding the encoder to map anomaly images to a generator-recognizable feature space for more stable reconstruction.
• Difference Analysis: Employing the U-Net network to perform pixel-level analysis on the original images, reconstructed images, and their residual images, thus detecting defect regions.
• Defect Synthesis: Designing a defect synthesis algorithm to generate various artificial defect images for training the U-Net network to improve its generalization ability.

Q: What are the main research findings and outcomes?

• Stable Background Reconstruction: By introducing the anomaly mapping method, it effectively eliminates defects of various sizes and obtains clear defect-free reconstructed images.
• Efficient Difference Analysis: The U-Net network effectively analyzes reconstruction differences and suppresses noise in normal regions, thus achieving more accurate defect detection.
• No Real Defect Samples Required: This method does not require real defect samples for training, making it suitable for surface defect detection problems where it is difficult to collect defect samples.
• Strong Generalization Ability: Experimental results on various texture image datasets and industrial production lines demonstrate the effectiveness and versatility of this method.

Q: What are the current limitations of this research?

• Limited Performance on Object Image Detection: The method has limited performance when detecting object images with a large number of unique local structures, as the generator and encoder have difficulty learning complex distributions of image patches at different locations.
• Long Training Time: The method requires a long training time, but the training process can be performed offline and does not affect the detection process.