DAUNet: direction-aware U-Net enhanced for FPC surface defect segmentation

Abstract

To support pixel-level surface defect segmentation for flexible printed circuits, this paper releases a synthetic segmentation dataset with fine-grained annotations for three typical defect categories, namely crush, stain, and scratch, to facilitate model training and evaluation. To address the challenges posed by large variations in defect scale, elongated defect shapes, and unclear boundaries, we further propose direction-aware UNet (DAUNet), a U-Net-based semantic segmentation network. The network introduces an inverted bottleneck cross-convolution module in both the encoder and decoder stages. By combining directional non-square convolution kernels, this module enlarges the receptive field and strengthens the modeling of directional patterns in linear scratches. In addition, a CCBAM module, which consists of channel attention and cross-shaped spatial attention, is embedded into the skip connections to improve selective multi-scale feature representation and fine-grained boundary modeling. Experimental results show that DAUNet achieves an mIoU of 77.15% on the synthetic dataset and only 2 false negatives and 3 false positives on 100 real production-line images, outperforming several mainstream semantic segmentation models overall.