CrackNeXt: An efficient local-to-context hybrid network for concrete surface crack classification in construction material inspection

Concrete cracking is a major indicator of material deterioration and serviceability loss, yet reliable image-based inspection remains difficult because crack morphology is often affected by surface texture, illumination, stains, shadows, construction joints, and other crack-like background patterns. This study proposes CrackNeXt, a compact local-to-context hybrid network designed for automated concrete surface crack classification. The architecture assigns different roles to different network stages: early ConvFormer-based stages preserve fine local crack cues such as thin discontinuities and texture breaks, while later Swin Transformer-based stages introduce contextual reasoning to reduce confusion with visually similar non-crack artifacts. CrackNeXt was evaluated on two complementary benchmarks. The METU dataset was used for binary crack classification under relatively controlled imaging conditions, whereas SDNET2018 was used for a more demanding six-class task that jointly models structural component and crack status across walls, decks, and pavements. Under a unified benchmark protocol against nineteen reference architectures, CrackNeXt achieved 99.90% accuracy on METU and 90.67% accuracy with a 90.29% macro-F1 score on SDNET2018. The model used 17.22 million parameters, required 2.66 GMACs, and achieved approximately 7 ms inference latency under the reported workstation setting. Grad-CAM visualizations indicated that high-response regions generally overlapped with crack-relevant image structures, supporting the qualitative plausibility of the learned representations. These findings suggest that staged local-to-context modeling can improve concrete crack classification under heterogeneous surface conditions.