Unpaired Image Denoising and Fusion with Adaptive Multi- Branch Task UNet for Semiconductor Packaging Defect Recognition

In advanced semiconductor packaging, products often exhibit high unit cost, complex structures, and stringent precision requirements. Inspection systems such as Scanning Acoustic Microscopy (SAM) are employed to detect internal defects, particularly within the Epoxy Molding Compound (EMC) layer. However, ultrasound images are susceptible to various noise sources that degrade image quality and hinder defect identification. This study proposes a self-supervised image denoising framework called Multi-Branch Task U-Net (MBT-UNet), which is a U-Net backbone with a multi-branch decoder. The model enables multi-task learning without requiring paired clean-noisy data or prior knowledge of noise characteristics. An empirical study of semiconductor packaging is conducted to validate the proposed MBT-UNet by comparing it with several benchmark methods (e.g., paired supervised learning approaches such as SC-UNet and RIDNet). The results show that MBT-UNet achieves competitive performance, as evaluated by the feature similarity (FSIM) and learned perceptual image patch similarity (LPIPS) metrics, and is comparable to supervised models despite being trained without paired data.