An Efficient End-to-End Multitask Network Architecture for Defect Inspection.

Area of Science:

• Computer Vision
• Materials Science
• Artificial Intelligence

Background:

• Automated steel surface defect detection is crucial but challenging due to defect complexity.
• Existing single-task networks struggle to meet comprehensive detection requirements.
• The limitations of current models necessitate a more robust and integrated approach.

Purpose of the Study:

• To develop an end-to-end multi-task network for improved steel surface defect detection.
• To address the challenge of varying defect scales using novel modules.
• To optimize network performance through strategic training methods.

Main Methods:

• An end-to-end multi-task network with one encoder and two decoders (object detection and semantic segmentation).
• A Depthwise Separable Atrous Spatial Pyramid Pooling module for dense multi-scale feature extraction.
• Residually Connected Depthwise Separable Atrous Convolutional Blocks for spatial information extraction.
• Investigated training strategies, including prioritizing segmentation and deep supervision.

Main Results:

• Achieved a mean Intersection over Union (mIOU) of 79.37% and mean Average Precision (mAP@0.5) of 78.38% on the NEU dataset.
• Demonstrated superior performance compared to existing models through comparative experiments.
• Reached a detection speed of 85.6 FPS on a single GPU, suitable for practical industrial use.

Conclusions:

• The proposed multi-task network effectively combines object detection and semantic segmentation for steel surface defect analysis.
• The novel modules and training strategies significantly enhance detection accuracy and efficiency.
• This approach offers a promising solution for real-world automated industrial inspection systems.