DSU-Net: A Defect Localization and Segmentation Algorithm for MEMS Sensors Using Skip Connections and Deformable Convolution

To further improve the accuracy of localization and segmentation of defects in microelectromechanical systems (MEMS) sensors, we present DSU-Net, a defect detection network built on a coding-decoding framework. DSU-Net builds upon UNet by incorporating dense skip connections, which strengthen the deep integration of high-level and low-level semantics. Additionally, deformable convolution is introduced to create a feature extraction module, boosting the model's capacity to capture defect characteristics. DSU-Net achieves excellent results on actual MEMS defect datasets. The experimental results indicate that our approach outperforms other leading methods in the field.