Line-Scan Defect Detection for Printed Laminated Films Using Multi-Template Matching and Residual-Driven Column-Band...

Machine vision enables efficient, non-contact industrial quality inspection. Online inspection of printed laminated films is challenging because complex graphics and transport-induced stretching cause appearance variations and spatially nonuniform mismatches. Global registration cannot fully correct these mismatches, whereas unrestricted dense deformation may absorb small defects. We propose an interpretable system based on a qualified multi-template library and residual-driven column-band compensation. Low-resolution retrieval selects geometrically compatible references; intermediate-resolution ORB-RANSAC registration and original-resolution warping then establish global correspondence. Compensation is activated only in residual-selected bands and constrained to the web-motion direction. Template-tolerant multi-scale Frangi comparison and component-level multi-template voting generate the final defect mask. The independent test set comprised 1500 production-line images: 200 defect-bearing images containing 823 annotated defect instances and 1300 defect-free images. Under the defect-instance-level evaluation, the system produced 765 true positives, 58 false negatives, and 31 false-positive predicted components, corresponding to 92.95% recall and 96.11% precision. The mean processing time was 357.06 ms/image. The results indicate that the method can balance residual-mismatch suppression and defect preservation under the evaluated production conditions.