Multiple Defects Inspection of Dam Spillway Surface Using Deep Learning and 3D Reconstruction Techniques

After a lengthy period of scouring, the reinforced concrete surface dam spillway (i.e., drift spillways and flood discharge spillways) will suffer from deterioration damage. Regular manual inspection is time-consuming dangerous. This paper presents robotic solution to detect automatically, count defect instance numbers, reconstruct by incorporating deep learning method with visual 3D reconstruction method. The lack real dataset incomplete registration minor defects on mesh model in fusion step are two challenges addressed paper. We created multi-class semantic segmentation 1711 images (with resolutions 848 × 480 1280 720 pixels) acquired wall-climbing robot, including cracks, erosion, spots, patched areas, power safety cable. Then, architecture U-net modified pixel-adaptive convolution (PAC) conditional random field (CRF) segment different scales defects, trained, validated, tested using this dataset. recovery instances flow sidewall facilitated keyframe back-projection By generating an adjacency matrix within class, intersection over union (IoU) voxels calculated fuse multiple instances. Our achieves average IoU 60% for five class. For model’s statistics, our accurate statistics area erosion environment 200 m2, absolute error number spots cracks has reduced original 13.5 3.5.