Copper frame diameters leave the caliper behind: zero-shot segmentation plus adaptive edges hold five new frames to 2.4–6.3 µm
Synopsis
Diameter inspection of CuCrZr frames for the Beijing Huairou High Energy Photon Source moves to a fully automatic vision pipeline: SAM3 extracts the target region without training, SAM-Canny adaptively picks the dilation kernel and Canny thresholds, and a WOA-RBF network compensates illumination-induced bias from grayscale features of different material regions. On five unseen frames the algorithm held errors to 2.4–6.3 µm, while manual micrometer measurement reached 32.1 µm.
Interpretation
Diameter inspection of CuCrZr frames shifts from manual contact measurement to a fully automatic vision pipeline whose readings no longer swing with the operator and the lighting. Such frames were previously measured mainly by manual operation of contact or non-contact instruments, which suits small-batch inspection but is inefficient and error-prone under the production demands of the photon source. On five previously unseen CuCrZr frames, the algorithm's absolute errors fell between 2.4 and 6.3 µm, whereas manual measurement with a Mitutoyo digital micrometer, averaged over ten readings per frame, spanned 1.3 to 32.1 µm.
Edge extraction no longer depends on hand-tuned parameters, because the dilation kernel size and the Canny double thresholds are both set from the image's own gradient statistics. Conventional Canny edge detection requires its high and low thresholds to be tuned for different imaging conditions, and existing adaptive variants either assume strong foreground-background separation, still need manually chosen percentiles, or carry high computational cost. Across 284 CuCrZr frame images spanning low to high illumination, SAM-Canny reached a maximum error of 18.15 µm, 35.8% below the 28.31 µm of fixed-threshold Canny, and raised the within-15-µm tolerance rate from 17.1% to 92.9%.
Illumination-induced systematic diameter bias is predicted from multi-material grayscale features instead of a single illumination reading. Earlier compensation strategies typically built a one-dimensional model linking illumination intensity to measurement error, which cannot separate grayscale changes caused by material reflectance from those caused by lighting. Using the mean grayscale of five SAM-segmented regions, the standard gauge block, the frame's outer, middle and inner rings, and the backplane, WOA-RBF reached an 82.3% within-tolerance rate at 10 µm, against 60.3% for one-dimensional least squares and 60.6% for GA-LSSVM.
Perspective
The pipeline is validated on diameter measurement of CuCrZr frames: training and error compensation use 284 images spanning different illumination intensities, generalization testing uses five new frames, and reference values come from a laser confocal microscope. The measurement platform is a 1920×1080-pixel CCD camera with a ring light source, calibrated with a 7×7 circular target. The authors propose future work to reduce dependence on manually specified prompts and to improve processing speed for real-time deployment.
SAM-Canny's standard deviation is 5.24 µm, slightly above the 4.52 µm of fixed-threshold Canny, which the authors treat as an acceptable trade-off given the accuracy and reliability gains. The compensation gain is measured at a 10 µm tolerance, and the model weights are fixed from the training experiment. The authors note the system still depends on manually specified prompts and that processing speed does not yet meet real-time deployment.