Trained CNN-SVM model weights — UAS flat-roof defect detection (3 seeds)

Trained model weights for the proposed tile-based CNN-SVM in "A Tile-Based CNN-SVM Architecture for Autonomous UAS Inspection of Flat Roofs." Three Keras models (best-validation weights) trained with random seeds 42, 43, and 44 on the leakage-free, photo-level split of the accompanying dataset; together they reproduce the reported mean test accuracy of 94.4% (95% CI ±0.4%). The model is a convolutional feature extractor with a linear squared-hinge (L2-SVM) output head; Defect = class 0 = positive. Load with keras.models.load_model(path, compile=False) — the training objective uses a custom squared-hinge loss and accuracy metric defined in run_experiments.py in the code repository.