SpecMamba-AD: A Fast and Accurate Framework for Micro-Scale Defect Detection in Industrial PU Suction Cups

Industrial pick-and-place automation requires rapid inspection of PU suction cups (vacuum cups) whose defects are subtle, low-contrast, and expensive to annotate at the pixel level. To address the combined challenges of normal-only training and shop-floor latency constraints, we propose SpecMamba-AD, an unsupervised teacher–student knowledge transfer framework for fine-grained defect localization in high-resolution industrial images. SpecMamba-AD pairs a frozen ImageNet-pretrained CNN teacher with a lightweight hybrid CNN–Mamba student. A blur-pooled, anti-aliasing CNN stem mitigates resolution-induced aliasing and token explosion, while Mamba blocks capture long-range structural consistency under limited memory. Training follows a denoising feature regression paradigm: the student receives synthetically corrupted inputs and is optimized to regress the teacher’s clean multi-level features, producing a discrepancy-based anomaly map at inference. To enhance sensitivity to micro-defects dominated by high-frequency cues, we further introduce a log-magnitude spectral loss that constrains teacher–student alignment in the frequency domain. Extensive experiments on PU suction cups data demonstrate that SpecMamba-AD achieves strong detection and localization performance under a strict normal-only protocol, and remains robust under realistic shifts in product line, illumination, and resolution supporting practical deployment for real-time quality assurance and stable grasp reliability.