Efficient Large-Kernel CNN with Cross-Scale-Attention Feature Fusion for AFP Defect Segmentation

Defect segmentation in the automated fiber placement (AFP) process demands both precise localization of sparse anomalies and efficient inference under resource constraints. To meet these requirements, we propose a novel segmentation network that integrates three key modules within a ConvNeXt-UPerNet framework. First, the proposed PeriShaper module replaces standard convolutions in the encoder with parameter-shared large-kernel branches, enabling shape-aware feature aggregation and expanded receptive fields at logarithmic parameter cost. Second, the plug-and-play cross-scale-attention feature fusion (CSAF) module employs unified-dimension cross-attention and multi-scale positional embeddings to deeply align and exchange information across feature pyramid levels. Third, the decoder is enhanced with a spatial convolution pyramid (SCP) and residual channel-attention (ResSE) blocks to recover fine-grained details. Evaluated on the AFP dataset of 460 pixel-annotated defect images, our model achieves a Dice coefficient of 0.5910 and IoU of 0.4752 (over 5% improvement compared with the baseline method), while inferencing at 39.13 FPS. Ablation studies confirm that each component contributes positively to performance. The proposed method balances segmentation quality and parameter efficiency, making it a promising solution for real-time AFP inspection systems.