With the rapid proliferation of tunnel and underground engineering projects, the safety assessment and routine maintenance of lining structures have become increasingly critical. Failure to timely identify and remediate common defects, such as cracks, segment damage, spalling of the secondary lining, and water leakage, can severely compromise the durability and operational safety of tunnel structures. Traditional manual inspection methods suffer from low efficiency and high subjectivity. Meanwhile, existing deep learning-based detection models frequently encounter challenges in complex tunnel environments, including uneven illumination, background interference, and constraints regarding deployment on mobile devices.

Participants are required to construct a lightweight, high-precision semantic segmentation model based on the provided tunnel lining image dataset. The model must be capable of performing pixel-level classification for multiple typical defect categories, including cracks, spalling of the secondary lining, segment damage, and various types of water leakage (leakage without discoloration/sediment, leakage with moss, and leakage with white crystallization). While ensuring detection accuracy, the model must also demonstrate fast inference speed and a compact model size to meet the deployment requirements of mobile or embedded devices.

•  
  Dataset Scale: The organizing committee provides 1,100 high-resolution images collected from real-world tunnel environments as training and validation sets for participants to download.
•  
  Data Characteristics: The dataset authentically reflects complex field conditions, featuring uneven illumination, significant background interference, and diverse surface textures. The images are provided in JPG and PNG formats. The annotations are in JSON format generated via LabelMe.
•  
  On-site Testing: The final evaluation will be conducted on-site utilizing a private test set. Participants are required to bring their trained models, generate inference results at the competition venue, and submit them immediately.

Participating teams are required to submit the following four deliverables:

•  
  Complete Code Package: Includes data preprocessing, model definition, and inference scripts.
•  
  Trained Model Weights: .pth, .onnx, .pb, or other standard formats.
•  
  Technical Report: Please submit a technical report in both PDF and Word formats. The content should cover model architecture, training strategy, and experimental results and analysis based on the provided 1,100 annotated images. All content must be written in English.
•  
  Inference Results (Submitted On-site): Instead of an execution script, participants are required to submit the inference result files generated during the on-site test. Participants must bring their trained models and generate the inference results on the designated test set provided at the competition venue. JSON files contain predicted shape information and class labels. The generated JSON files must be submitted immediately after the inference process.

The final score is comprehensively determined by detection accuracy, model size, and expert evaluation.

•  
  Formula: Score = Accuracy Score × Size Coefficient + 0.3 × ExpertScore.
•  
  Accuracy Score: mIoU across 7 classes on the test set × 100.
•  
  Size Coefficient: 1.1 for models ≤50MB; 1.0 for 50MB-100MB; 0.9 for models ≥100MB.
•  
  ExpertScore: Calculated as the arithmetic mean of scores awarded by review experts based on the technical documentation and on-site defense.