Image-based intelligent detection of typical defects of complex subway tunnel surface

Abstract

We propose a new hybrid model algorithm that addresses the challenges of recognizing multiple objects. Specifically, objects are recognized from massive, large-scale, and complex tunnel images without repeated parameter adjustments and high-cost annotation datasets. Our algorithm utilizes a multi-scale, fusion-based encoder–decoder segmentation model to classify objects from high-resolution images of the tunnel surfaces. To enhance the accuracy of crack identification from complex backgrounds, we incorporate the Expanded Threshold Search (ETS) algorithm and the Local Window Extraction (LWE) algorithm. The acquisition device and the algorithm, implementing the multi-object dataset, have successfully tested, whereby it recognizes five objects and attains the highest Intersection over Union (39.3% for the crack object, 65.6% for the leakage object, and 75.7% for the rest).