Automated visual inspection of civil infrastructure has become a central application field for deep learning, yet the gap between performance on curated datasets and robustness under real-world conditions remains substantial. This contribution formalizes and extends an internal benchmarking study of a ResNet50-based classifier (Aigis), EfficientNet-B0 and B2, and an adapted YOLOv8n-seg model for image-level classification of four concrete surface conditions: corrosion, crack, spalling, and normal structure. Critically, the YOLOv8n-seg adaptation employs a heuristic inference strategy: images with no positive detections are defaulted to “Normal Structure", while others are assigned the class of the highest-confidence instance. Using two datasets—a controlled set of 3, 608 labeled images and a more heterogeneous corpus of 21, 013 images collected via Roboflow—we analyse accuracy, per-class precision, recall, F1-score, ROC–AUC, and confusion patterns. EfficientNet-B2 with test-time augmentation achieves \(88.59\%\) accuracy on the original dataset, whereas the YOLOv8n-seg model, constrained by the misalignment between local detection and global categorization, collapses to \(42.11\%\) . When evaluated on the larger, more variable dataset, the ResNet50-based Aigis model drops from \(68.01\%\) to \(58.64\%\) accuracy, with spalling recall falling to \(10.4\%\) . These results are interpreted in the light of contemporary literature on crack detection, efficient convolutional backbones, and domain adaptation in structural health monitoring, and they motivate concrete research directions in multiclass benchmarking, model-task alignment, and generalization under dataset shift.