Industrial digital radiographic image denoising based on improved KBNet

Area of Science:

• Industrial Imaging
• Image Processing
• Artificial Intelligence

Background:

• Industrial digital radiography (DR) images are crucial for inspections but often degraded by noise, scatter, and cross-talk.
• Challenges in denoising include non-zero mean noise and neighborhood correlation loss in 1D array scanning.
• Existing methods struggle to effectively address low resolution and noise in industrial DR images.

Purpose of the Study:

• To enhance the denoising performance of industrial DR images.
• To address limitations of existing denoising techniques in industrial applications.
• To develop a computationally efficient and effective denoising model for industrial DR images.

Main Methods:

• Proposed an improved KBNet (iKBNet) incorporating lightweight modifications.
• Integrated the Convolutional Block Attention Module (CBAM) to reduce network parameters.
• Utilized Structural Similarity Index (SSIM) loss within a composite loss function for improved denoising.

Main Results:

• The iKBNet demonstrated superior denoising performance compared to BM3D, ResNet, DnCNN, and the original KBNet.
• Achieved higher image restoration quality metrics.
• Produced satisfactory results in practical applications with low-resolution transmission images.

Conclusions:

• The iKBNet effectively minimizes computational cost and enhances processing speed.
• The model achieves superior denoising results, improving the quality of industrial DR images.
• iKBNet offers a promising solution for processing noisy digital radiographic images in industrial settings.