Inline Visual Defect Screening under Few Normal Samples and Illumination Shift: A Measurement Study of Memory-Augmented State-Space Detectors on Edge Hardware

Abstract

Inline visual screening stations are chosen under three constraints at once: a handful of confirmed good parts after a product changeover, illumination that drifts between shifts, and a hard per-part latency budget at line speed. We build a single-camera station model (5 MP, 60 × 50 mm field of view, 12 parts/s, 83 ms budget, Jetson Orin Nano 15 W) and measure seven detectors—PaDiM, PatchCore, RD4AD, DRAEM, EfficientAD, MambaAD, and a memory-augmented selective state-space detector (MS-AD)—on MVTec AD and VisA under k ∈ {1, 2, 4, 8, 16} normal images per class and three illumination conditions (800, 1,500, and 3,000 lx), using defect-level recall at 1% false rejection as the operating metric. From two normal images upward, MS-AD gives the best recall (0.879 on MVTec AD and 0.758 on VisA at k = 8 and 800 lx); with one normal part, a patch-memory method is preferable. Illumination drift costs 2.7–7.4 AUROC points—more than the choice of detector—and misses concentrate on low-contrast defects below 0.4% of the image area. On the edge module, the memory stage, not the state-space scan, dominates latency; 320 px INT8 input with M = 256 memory slots is the largest configuration inside the 83 ms budget (52.8 ms). Per-condition score normalisation restores a 1.3–2.4% false-rejection rate across conditions. A 120-part physical re-shoot anchors the simulated lighting conditions within 0.7 AUROC points. The protocol and its measurements transfer to single-camera stations with comparable optics and line speeds.