X-Enhanced ULite: Improving semantic segmentation for surface defects

In surface defect segmentation, achieving high accuracy while maintaining lightweight model characteristics is crucial for industrial applications. However, designing lightweight models that achieve high accuracy without compromising computational efficiency remains a significant challenge. In this work, we propose X-ULite, an enhanced version of the ULite framework, designed to improve both accuracy and efficiency in defect segmentation. The core innovation lies in the XConv module, which decouples main and anti-diagonal convolutions into independent depthwise branches, thereby preserving orientation-specific features that are crucial for representing cracks and scratches. Additionally, a new BottleNeck module integrates XConv with Axial Depthwise Convolution (AxialDW) and standard depthwise convolution, jointly modeling axial, diagonal, and local features to achieve a comprehensive perception of defect regions. Evaluated on industrial datasets, XULite achieves 85.98 % mIoU on steel surfaces (NEU-Seg), 75.07 % mIoU on Magnetic Tile Defect datasets, and 91.44 % mIoU on Mobile phone screen surface defect (MSD) datasets with only 0.97M parameters. The model maintains low computational complexity and parameters while demonstrating robust segmentation accuracy across diverse industrial scenarios.