Detecting apparent defects is fundamental for assessing the health of structures and provides guidance for preventive maintenance toward mitigating engineering hazards. Deep learning-driven computer vision methods have recently gained prominence in intelligent defect identification. However, the missed detection of small-scale targets and the unbalanced accuracy across defect categories are limitations of existing algorithms, owing to substantial variations in spatial scales. To overcome these limitations, this study introduces a detection approach that effectively captures deep network feature representations, mitigates the loss of semantic information for small targets, and maintains balanced recognition performance across multiple defect types.

A lining apparent defect detection network (LADDNet) that integrates an information-sharing backbone with an adaptive and hierarchically organized feature fusion mechanism is proposed. First, a three-branch collaborative feature extraction architecture is constructed by incorporating lightweight ShuffleNet and GhostNet modules into a CSPNet framework. This design facilitates complementary feature learning across branches, thereby enhancing the stability of gradient propagation. Thereafter, an attention-integrated multi-receptive field adaptive fusion (AMFAF) module is developed. This module employs parallel convolutions with diverse receptive fields to extract multi-level spatial information and combines them via attention-based weighting, allowing the network to automatically emphasize discriminative features associated with cracks, seepage regions, and spalling contours. An attention-based intra-scale feature interaction (AIFI) module is also introduced to enhance semantic consistency within individual feature scales by promoting effective cross-channel communication and suppressing redundant responses. Finally, an adaptively spatial feature fusion detection head (ASFF-Head) is incorporated to refine multi-scale feature aggregation, improve the localization precision, and reduce missed detections of small or low-contrast targets. By integrating these modules into a unified framework, the proposed network supports end-to-end training and inference.

LADDNet exhibits significant advantages in the overall detection performance and inference efficiency. On the validation set, the model achieves F1, mAP@0.5, and mAP@0.5: 0.95 scores of 0.831, 0.848, and 0.595, respectively. On the test set, the model attains an F1 score of 0.794, mAP@0.5 of 0.830, and mAP@0.5: 0.95 of 0.579. Compared with a range of representative detection models, LADDNet achieves consistent improvements in both the F1 score and mAP@0.5. In terms of inference efficiency, LADDNet achieves real-time performance with a per-image latency of 9.2 ms, only 14.07×10⁶ parameters, and 21.4×10⁹ FLOPs, delivering substantially faster inference than RT-DETR. Furthermore, when detecting defects in images containing handwritten markings or interference from auxiliary tunnel facilities, LADDNet continues to demonstrate strong robustness. For the identification of mesh cracks, water seepage, and spalling, the model demonstrates high confidence, low missed-detection rates, and precise localization.

The proposed LADDNet model affords markedly enhanced intelligent detection of diverse tunnel-lining defects by integrating information sharing, multi-receptive-field feature extraction, adaptive fusion strategies, and intra-scale interaction mechanisms. It delivers notable gains in accuracy, robustness, and computational efficiency, effectively overcoming long-standing challenges in multi-scale and multi-type defect recognition. These advances position LADDNet as a reliable visual perception module for automated tunnel inspection, structural condition evaluation, and long-term operational monitoring. Overall, the approach shows strong promise for real-world engineering applications and broad deployment in next-generation intelligent maintenance systems. Its versatility further underscores its value for future infrastructure management initiatives.