Correction of motion artifact in CL based on MAFusNet

Area of Science:

• Industrial Nondestructive Testing
• Image Processing
• Computed Laminography

Background:

• Computed laminography (CL) is a key technique for nondestructive testing of plate-like objects.
• Continuous movement during CL scanning enhances data acquisition efficiency but introduces motion artifacts.
• Existing methods struggle to effectively correct these motion artifacts without significant data loss.

Purpose of the Study:

• To develop a novel deep learning network for correcting motion artifacts in CL projection images.
• To improve the quality and efficiency of CL imaging in industrial applications.
• To leverage multi-angle information for enhanced deblurring capabilities.

Main Methods:

• A multi-angle fusion network (MAFusNet) was designed, incorporating multi-angle fusion and feature fusion modules.
• The network utilizes data from nearby projection images to improve deblurring.
• MAFusNet was trained on synthetic datasets and validated on realistic data.

Main Results:

• MAFusNet demonstrated significant improvement in correcting CL motion artifacts compared to conventional deblurring networks.
• The multi-angle fusion module enhanced the network's deblurring ability.
• The use of synthetic datasets proved effective for training, showing strong generalization to real-world data and reducing training costs.

Conclusions:

• MAFusNet effectively corrects motion artifacts in computed laminography, enhancing image quality and efficiency for industrial nondestructive testing.
• The network's architecture and synthetic data training approach offer a robust and cost-effective solution.
• This method shows great potential for advancing CL imaging techniques.