A research team at Southwest Jiaotong University, working with machine-vision integrator CENTECH-EG, has demonstrated an automated inspection cell that cuts the missed-detection rate for surface defects on coated cemented carbide CNC inserts from 21.4% to 11.8% relative to a baseline YOLOv8 detector. The work, published in Scientific Reports on 11 May 2026 (DOI: 10.1038/s41598-026-52293-1), targets a gap in the quality chain: insert inspection still leans heavily on manual visual judgment, and no public dataset existed for this part family.

The optics and mechanics

The cell occupies roughly a 2 m × 2 m footprint. A KUKA KR 6 R900-2 six-axis arm (±0.02 mm repeatability, 901 mm working radius) picks each insert from a tray, guided by a positioning camera, and presents it to the imaging chain: a 5120 × 5120 pixel camera behind a telecentric lens, yielding 2.5 µm per pixel over an 8.5 mm² field. Each square insert is imaged in four overlapping shots so no edge escapes coverage. Illumination comes from a ring LED combining bright-field and dark-field paths — bright-field renders chips and bubbles as dark features on a bright background, dark-field catches raised defects as bright scatter against darkness, and the 360° annular geometry suppresses the specular glare typical of polished carbide coatings.

The dataset

The authors built a 651-image dataset (522 training, 129 validation, 4:1 split) and, for the first time, systematized coated-carbide insert defects into seven classes: chipping, protrusion, notch, bubble, edge depression, color aberration, and material loss. Annotators used X-AnyLabeling 2.5.3 with COCO-format bounding boxes; two authors labeled independently, with a third adjudicating disagreements. The dataset shows heavy intra-class variance — "material loss" instances vary widely in size and shape — and inter-class similarity between "notch" and "edge depression." Scale spread is significant: chipping instances range from roughly 1000 × 500 down to 400 × 200 pixels; most notches sit near 300 × 100 pixels. The team intends to release the dataset publicly.

The model

Their detector, A2MD-YOLO, modifies YOLOv8 in two places. A P2 detection head taps shallower, higher-resolution feature maps for small-defect sensitivity — the standard P3/P4/P5 pyramid favors large receptive fields at the expense of local detail. A Convolutional Block Attention Module (CBAM), inserted at the end of the backbone before SPPF, recalibrates channel and spatial responses so the network concentrates on edge-adjacent regions where defects cluster.

Ablation results show a deliberate trade. Precision drops from 0.791 to 0.661, but recall climbs from 0.595 to 0.733, F3-score rises from 0.610 to 0.725, and mAP50-95 improves from 0.529 to 0.571. The authors chose β = 3 in the F-score explicitly because in industrial QA a missed defect carries more risk than a false alarm routed to manual recheck. Against YOLOv11, YOLOv12, and the 2025-released YOLO26, the plain YOLOv8 baseline itself outperformed the newer versions on precision and mAP50-95 on this dataset — a reminder that architecture generations do not automatically transfer to small industrial datasets. On the validation set of 220 defects, detection accuracy rose from 63.6% to 73.2%. The model detected protrusion and bubble classes perfectly; residual confusion occurred among visually similar categories.

Throughput and deployment

Measured in the deployed cell, inspection averages about 25 seconds per insert; a full 12 × 8 tray takes roughly 40 minutes. That number deserves context. A medium-sized manufacturer cited by the authors produces around 10 million inserts per year — about 5,000 per hour — and manual visual inspection takes roughly 2 seconds per insert. The system is therefore not an in-cycle online inspector. The authors position it as an offline, end-of-line automation stage: gripping, imaging, detection, and automatic sorting into good and defective trays, running 24/7 with an operator handling only tray placement and removal. Future work targets shorter arm travel paths to compress the cycle.

An uncertainty analysis adds rigor often missing from vendor claims. Across the 15 missed-detection images, at least one of three quality metrics — signal-to-noise ratio, mean gradient magnitude, or Laplacian variance — fell below the validation-set average, tying misses to imaging physics rather than model failure alone. Surface contamination on real production inserts introduces random reflections and scattering that degrade both false-negative and false-positive rates.

Against a Keyence VHX-X1 digital microscope (accurate but manual, operator-dependent) and Zeng et al.'s 2024 conveyor-based CBN insert system (fixed-angle imaging, four defect classes, no public data), the authors argue their cell offers multi-surface coverage, full automation, and a reusable dataset.

The open question for adopters: with 66% precision, one in three flags is a false alarm requiring human review, and 25 s per part limits the cell to batch and end-of-line roles. Whether the promised public dataset materializes — and whether recall gains hold across other insert geometries and materials — will determine if this approach scales beyond square coated carbide parts.