GLASS-FFT: A fast and accurate framework for micro-scale defect detection in linear guideway end seals

Micro-scale linear guideways require rapid inspection of face-seal gaskets whose defects are minute and low-contrast. High labeling costs and a 20[Formula: see text]s takt time limit the practicality of fully supervised CNNs. We propose GLASS-FFT-SA, which fuses dual-layer anomaly synthesis with spectrum-efficient convolution and a lightweight spectral attention block to amplify high-frequency cues. Local synthesis inserts defect textures into normal images to create strong anomalies, while global synthesis performs gradient ascent on the feature manifold to generate subtle, near-boundary cases, yielding abundant, diverse training data. Replacing large [Formula: see text] and [Formula: see text] spatial kernels in a ResNet-34 backbone with FFT-based convolutions reduces complexity and latency; a gated cross-attention mechanism triggers the pixel branch only when the image head’s anomaly score exceeds a learnable hard-sigmoid gate. Trained on 10,000 normal images and 8000 synthetic anomalies, GLASS-FFT-SA attains image-level AUROC 0.97, pixel-level AUROC 0.96, AUPRO 0.94 and 35 FPS on RTX 3090 — matching the precision of GLASS while running [Formula: see text] faster. It sustains AUROCs [Formula: see text] across shifts in product type, illumination and resolution, and outperforms unsupervised baselines under identical preprocessing, input resolution and fixed thresholds. These findings show that spectrum-efficient convolution plus targeted anomaly synthesis yields a favorable accuracy–throughput trade-off for fine-grained defect inspection.