Defect Detection of Industry Wood Veneer Based on NAS and Multi-Channel Mask R-CNN

Jiahao Shi, Zhenye Li, Tingting Zhu, Dongyi Wang, Chao NiSensors 2020 · DOI

TL;DR

This paper introduces a novel wood veneer defect detection system that significantly improves both accuracy and speed. By combining a Neural Architecture Search (NAS)-designed "glance" network for rapid initial screening with a multi-channel Mask R-CNN for detailed analysis, the system achieves state-of-the-art performance, making it suitable for real-time industrial applications.

What Problem Does This Paper Address?

The wood veneer industry relies heavily on visual inspection to identify defects like knots and cracks, which can significantly impact the commercial value of the final product. While existing defect detection methods often prioritize accuracy, they frequently fall short in terms of speed, making them unsuitable for the high-throughput demands of modern manufacturing. This paper addresses this critical gap by proposing a system that balances high detection accuracy with the speed required for online, real-time applications.

Methods

The core innovation lies in a two-stage detection pipeline:

  • Glance Network (NAS-based): This network acts as a fast pre-filter. It's designed using Neural Architecture Search (NAS) technology, specifically optimized for speed (minimizing Floating-Point Operations - FLOPs) while maintaining a high detection rate. NAS is employed to automatically discover the most efficient network architecture for quickly identifying potentially defective images from a large batch of regular ones. This significantly reduces the computational load on the subsequent, more complex detection module.
  • Multi-Channel Mask R-CNN: For images flagged as potentially defective by the glance network, a multi-channel Mask R-CNN is used for detailed analysis. This includes:
  • Feature Extraction: The "first impression" features from the glance network are combined with the original image features.
  • Genetic Algorithm for Feature Selection: A genetic algorithm is used to intelligently select the most informative channels (features) from the glance network's output to feed into the Mask R-CNN, further optimizing accuracy.
  • Mask R-CNN: This established architecture is used for precise defect localization and segmentation, outputting the shape and type of the defect.

Datasets

The study utilized a custom dataset of 2838 wood veneer images collected using specialized acquisition equipment. This dataset was carefully curated, with images being cropped, augmented (rotation, magnification, mirroring), and then labeled by experienced workers. The dataset comprises:

  • 612 regular (background) images.
  • 2226 defective images, categorized into:
  • 846 live knot pictures
    • 760 dead knot pictures
    • 620 crack pictures

The area of defects in these images ranged from 12.086 to 313.238 mm².

Key Insights and Ideas

  • The Power of Pre-filtering: The "glance network" concept is crucial. By quickly discarding the vast majority of non-defective images, the system dramatically reduces the computational burden on the more resource-intensive Mask R-CNN.
  • NAS for Speed Optimization: Applying NAS not just for accuracy but explicitly for speed (FLOPs minimization) is a key contribution, enabling the glance network to be truly efficient for industrial use.
  • Synergistic Feature Combination: Merging features from the initial glance network with the input to the Mask R-CNN provides richer information, leading to improved detection accuracy.
  • Genetic Algorithm for Channel Optimization: Using a genetic algorithm to select the most relevant features for the Mask R-CNN input is an effective way to fine-tune performance without manual trial-and-error.
  • Adjustable Confidence: The introduction of an adjustable confidence rate for the glance network allows for a trade-off between detection rate and false alarm rate, with the understanding that the subsequent Mask R-CNN can handle false positives.

Key Results

The proposed GM-Mask R-CNN model demonstrated impressive performance:

  • Overall Classification Accuracy (OCA): 98.70%
  • Mean Average Precision (MAP): 95.31% ± 4.5
  • Inference Time: Only 2.5 seconds to process a batch of 100 images (50 defective, 50 regular).

Compared to traditional Mask R-CNN and other methods like SegNet and FCN, the proposed approach achieved higher accuracy and significantly reduced inference time. For instance, it was over 8 times faster than SegNet. Individual defect detection accuracies were also high, with cracks reaching 100% and live knots at 96.74%.

Limitations and Future Work

  • NAS Training Time: The authors acknowledge that the NAS process for designing the glance network can be time-consuming, especially as the search space increases. They plan to explore more efficient search strategies.
  • Joint Optimization: The current model uses separate optimization for the glance network and the Mask R-CNN. Future work will focus on developing a joint optimization approach to reduce overall complexity and potentially improve performance.
  • Expanding Defect Types: The current model focuses on three defect types. Future research aims to expand its capability to detect a wider variety of defects.
  • Defect Area Calculation: The authors intend to extend the model to calculate the area of detected defects, which would provide valuable quantitative data for industrial applications.

Summary updated

BibTeX
@article{shi2020defect,
  author = {Jiahao Shi and Zhenye Li and Tingting Zhu and Dongyi Wang and Chao Ni},
  title = {{Defect Detection of Industry Wood Veneer Based on NAS and Multi-Channel Mask R-CNN}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20164398},
}
Figure