A lightweight multi scale fusion network for IGBT ultrasonic tomography image segmentation

Area of Science:

• Power Electronics
• Non-Destructive Testing
• Image Analysis

Background:

• Insulated Gate Bipolar Transistors (IGBTs) are critical power semiconductor devices where internal structural integrity dictates performance and reliability.
• Semantic segmentation of IGBT ultrasonic tomographic images is challenging due to high-density noise and target warping-induced visual distortions.

Purpose of the Study:

• To develop a robust and efficient method for semantic segmentation of IGBT ultrasonic tomographic images.
• To address the limitations of existing methods in handling noise and distortion in these specific images.

Main Methods:

• Construction of a dedicated IGBT Ultrasonic Tomography (IUT) dataset using Scanning Acoustic Microscopy (SAM).
• Proposal of a lightweight Multi-Scale Fusion Network (LMFNet) with a U-shaped encoder-decoder architecture and inverted residual blocks.
• Introduction of Context Feature Fusion (CFF) and Multi-Scale Perception Aggregation (MPA) modules for enhanced feature integration.

Main Results:

• LMFNet achieved superior segmentation accuracy on the IUT dataset compared to existing methods.
• The proposed network demonstrated significant improvements in model lightweighting performance.
• Experimental validation confirmed the effectiveness of LMFNet in handling noise and distortion.

Conclusions:

• LMFNet offers an effective solution for accurate and efficient semantic segmentation of IGBT ultrasonic tomographic images.
• The developed dataset and network architecture contribute to advancements in power semiconductor non-destructive testing and analysis.
• The lightweight design of LMFNet facilitates practical implementation in industrial inspection scenarios.