AFCC: Adaptive Fusion of Local Contour Descriptors and Hierarchical Convolutional Features for Weld Formation Classification on Curved Workpieces

Abstract Reliable visual inspection of welds on curved workpieces remains challenging because non-planar geometry, viewpoint variation, and non-uniform reflection cause the same weld structure to appear inconsistently across images. This paper proposes AFCC, an adaptive fusion framework that combines local contour descriptors with hierarchical convolutional features to classify weld formation as normal, excessive/distorted, or insufficient. Hu moment descriptors computed over sliding windows of an edge-enhanced image characterize localized bead geometry, while five groups of intermediate EfficientNet-B0 embeddings represent edge, texture, structural, texture-and-saturation, and high-level semantic information. Rather than concatenating the streams or assigning equal weights, AFCC weights each stream according to its chance-corrected balanced accuracy on the validation folds of five-fold cross-validation. The framework was evaluated on 1,586 weld images acquired at-line in an industrial metal-rack production facility, including a held-out test set of 477 images. With a Siamese neural network classifier, AFCC achieved 95.39% accuracy, 94.32% macro recall, and a 92.31% macro F1 score. Its macro F1 score exceeded those of a fine-tuned EfficientNet-B0 baseline and a representative deep-feature fusion method by 13.95 and 7.59 percentage points, respectively. AFCC missed only 4 of the 120 defective test welds, compared with 14 and 8 for these baselines, and its three-class output can route each weld to grinding, supplementary welding, or pass-through. Ablation results further indicated that the contour and convolutional representations are complementary and that reliability-guided weighting improves defect detection beyond equal weighting.