Residual Intelligence: Deep Residual Learning for Image Recognition in Construction – Applications and Future Directions

Rapid growth in machine learning advancement has positioned computer vision as a linchpin of data‑driven construction practice, yet the sector’s reliance on deep residual networks remains poorly charted. To clarify the state of the art, we conducted a PRISMA guided bibliometric review of Scopus records published between 2016 and March 2025. A tightly scoped query linking residual‑learning terms with image‑recognition tasks and built‑environment contexts returned 292 items; after language and document‑type screening, 258 publications were analysed. Co‑citation, keyword and country overlays were generated in VOSviewer and interpreted against domain knowledge. Results reveal a six‑fold rise in annual outputs since 2019, with China, the United Kingdom and India accounting for half of all papers. Intellectual structure converges on three mature pillars (semantic segmentation, object detection and image classification) and two emergent strands: resource‑aware/federated learning and sustainability‑oriented applications. Transformer enhanced detectors, edge‑optimised residual hybrids and diffusion‑based data augmentation are identified as recent inflexion points. Conversely, limited cross‑domain transfer studies, scarce explainability protocols and an absence of carbon footprint assessments expose critical research gaps. The review thus furnishes a data‑driven roadmap for shifting residual intelligence from isolated prototypes to trustworthy, resource‑efficient and globally transferable vision systems capable of underpinning autonomous, safe and sustainable construction operations.