Defect detection in rubber pump tubing using edge fusion and adaptive learning

Abstract

Rubber pump tubing governs the volumetric accuracy of peristaltic infusion pumps, so deviations in inner diameter or wall thickness can alter the delivered dose. Automated optical inspection of such tubing is constrained by two coupled factors: the samples themselves vary geometrically, which injects nuisance variation into the input distribution, and the diagnostic evidence is carried by thin, low-contrast edges that repeated downsampling in a convolutional network attenuates. This study addresses both factors within a single pipeline and attributes their contributions separately. A shadow-line detection procedure first recovers the inner and outer boundaries of the two tube walls from a vertical darkness profile, computes layer-wise inner diameter, outer diameter and left/right wall thickness, and uses the resulting asymmetry statistics as an explicit and interpretable data-conditioning stage. A VGG19 backbone is then augmented with a Residual Edge Fusion Unit, in which a fixed Sobel operator yields a gradient magnitude that a two-parameter learnable gate converts into a soft edge response that is re-injected into the backbone features through a residual path rather than concatenated as additional channels, and with a parameter-free adaptive average pooling head that removes the fixed input-size constraint of the original classifier. Transfer learning is applied in a partial-freeze form to accommodate the limited sample size: the first four convolutional blocks of the ImageNet-pretrained backbone are held fixed for the whole run, the deepest convolutional block is fine-tuned, the fusion unit is trained from scratch, and within the classifier only the final fully connected layer is replaced and trained while the preceding layers stay frozen. Experiments were conducted on 3000 annotated rubber pump tubing images acquired on a line-scan inspection rig. On a test partition of 600 images held out at the group level, the full pipeline attains a positive-class F1 of 0.9093 and an accuracy of 0.8733, against 0.8897 and 0.8483 for the same network without geometric screening and 0.8317 and 0.7700 for the unmodified VGG19 baseline; over five random seeds the F1 of the full pipeline is 0.9094 with a standard deviation of 0.0010, its bootstrap 95 per cent confidence interval is [0.8881, 0.9286], and a paired McNemar test separates it from the unscreened variant at p = 0.0201.