Generalized denoising network LGCT-Net for various types of ESPI wrapped phase patterns

Area of Science:

• Optical Metrology
• Image Processing
• Artificial Intelligence

Background:

• Electronic Speckle Pattern Interferometry (ESPI) is crucial for non-destructive testing and deformation analysis.
• Speckle noise in ESPI phase patterns degrades measurement accuracy and hinders analysis.
• Existing denoising methods often struggle with diverse phase pattern characteristics like varying densities and discontinuities.

Purpose of the Study:

• To propose a generalized network, LGCT-Net, for effective denoising of various ESPI wrapped phase patterns.
• To introduce the Local-Global Channel Transformer (LGCT) module, integrating local and global feature extraction.
• To develop a comprehensive dataset for training and evaluating ESPI phase pattern denoising.

Main Methods:

• The proposed LGCT module combines Dilated-Group Convolution (DGC), Contextual Transformer (CoT), and Efficient Channel Attention (ECA) blocks.
• LGCT-Net features a dense connection architecture interleaving LGCT modules and Conv+BN+ReLU layers.
• A diverse dataset of simulated and experimental ESPI phase patterns was created for training and validation.

Main Results:

• LGCT-Net successfully denoises low-density, medium-density, high-density, variable-density, and discontinuous ESPI phase patterns without preprocessing.
• The method significantly reduces speckle noise, enhances fine details, and preserves structural integrity.
• Quantitative and qualitative comparisons show LGCT-Net outperforms established methods like PEARLS, HDCNN, ADCNN, and DBDNet.

Conclusions:

• LGCT-Net provides a robust and effective solution for denoising complex ESPI wrapped phase patterns.
• The proposed LGCT module's ability to leverage both local and global information is key to its superior performance.
• The method demonstrates successful application in dynamic measurements and subsequent phase unwrapping of nuclear graphite ESPI data.