Answers are grounded in this paper and cite their sources.
Accurate recognition of aluminum profiles from images of their cross-sections is important for industrial identification and inventory-related applications. However, developing robust recognition models is challenging because aluminum-profile classes can exhibit highly similar geometries, while sufficiently large and consistently annotated real-image datasets are difficult to obtain. This study presents a proof-of-concept synthetic-to-real framework for the cascaded recognition of aluminum-profile cross-sections. A geometry-preserving data-generation algorithm transforms CAD representations of the profiles into synthetic training images and introduces controlled variations in scale, position, background, clutter, and appearance. The resulting data are used to train cascaded detector–classifier configurations, which are evaluated in terms of recognition performance and computational efficiency. The experiments show that models trained exclusively on synthetic data can localize aluminum profiles in real images but may fail to classify them correctly because of differences between synthetic and real visual characteristics. After few-shot fine-tuning on 21 real photographs with the shallow layers frozen, the selected YOLO26n + YOLO26n-cls cascade achieved an end-to-end accuracy of 68.2% (95% CI: 47.3–83.6%) on an independent final evaluation set of 21 photographs containing 22 objects; the YOLOv8s + YOLO26n-cls cascade achieved 77.3%. A separate pre-adaptation diagnostic on the original 21-record image-level manifest produced no correct classifications. Because that diagnostic used a different counting rule and an IoU threshold of 0.25, it is interpreted as evidence of direct-transfer failure rather than as a matched numerical baseline for the post-adaptation results. The selected cascade required approximately 0.6 s per classified object in a CPU TensorFlow Lite benchmark. With 22 test objects the intervals are wide, and the classifier comparison is confounded by differences in adaptation recipes. The findings also highlight the importance of evaluating localization and classification separately when analyzing synthetic-to-real recognition systems. To demonstrate its practical applicability, the resulting recognition pipeline is integrated into Signal Profiler, a mobile system that performs image acquisition, profile localization, cross-section classification, and result storage. Since the evaluation contains only 21 photographs (22 object instances), it does not establish generalization across profiles, cameras, or acquisition conditions.
No takes yet. Share an insight, caveat, or question.
Kafestidis et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: