Fast-SDNet: an extremely lightweight and accurate surface defect segmentation network

Automated surface defect segmentation is critical in industry, but accurate detection is challenging due to diverse defect types, large-scale variations, and high background similarity. Limited on-site computational resources further restrict algorithm deployment. To address this, we propose an extremely lightweight and accurate surface defect segmentation model named Fast-SDNet. The structural optimization of Fast-SDNet encompasses the following components: 1) Dense shuffle group convolution module, an extremely efficient feature fusion method for combining shallow and deep features. 2) Global-local attention module, providing long-distance information capture and local information focus to enhance semantic interaction. 3) Edge-enhanced loss, strengthening the model’s learning of defect edge regions to improve fine-grained segmentation of edge regions. Experimental results on three different industrial surface defect datasets demonstrate significant advantages in mIoU, model parameters, and FLOPs. Specifically, the proposed method achieves mIoU scores of 65.8%, 73.9% and 76.3% on the KolektorSDD2, RSDD1 and NEU-SEG datasets, with Params and FLOPs only accounting for 0.18% and 0.24% of U-Net, respectively. Additionally, Fast-SDNet achieves an extremely low inference latency on CPU-only devices, enabling real-time inspection with low-cost hardware and substantially lowering the deployment barrier in industrial settings. Code is available at: https://github.com/Wanglaoban3/Fast-SDNet.