Pixel-level damage detection for concrete spalling and rebar corrosion based on U-net semantic segmentation

Although computer vision and image processing techniques have been introduced to perform pixel-level damage detection from images, the existed studies mainly investigated cracks. However, concrete spalling and rebar corrosion regions are always irregular and lathy. Effective recognition models of these damage types still lack, especially in the real-world situations with complicated background disturbances. This study proposed a novel pixel-level damage detection method for concrete spalling and rebar corrosion based on U-net semantic segmentation. The proposed U-net architecture consists of three encoder and decoder stages, which are connected by a bridge section. A series of convolutional, batch normalization and ReLU layers are arranged in sequence, and max pooling, dropout, transposed convolutional and depth concatenation layers are specifically designed. A weighted pixel segmentation layer is designed to solve the unbalanced data problem. The total number of original images is 1763 and their dimensions are 512 × 512 × 3. 85% of images are used for training, 10% for validation, and the rest 5% for test. Data augmentation including random rotation, flipping, shearing, translation, and scaling is adopted in the training process. The final training accuracy reaches 85.85% after all the 1350 training iterations of 50 epochs. The average recognition precision and recall for background, concrete spalling and rebar corrosion are0.7662 and 0.7697, respectively. The global accuracy, mean accuracy, mean IoU, weighted IoU reaches 0.7815, 0.7697, 0.3747, and 0.7251, respectively. The category-wise IoU of background, concrete spalling and rebar corrosion regions reaches 0.7842, 0.2629, and 0.0770, respectively. It shows the rebar corrosion has the lowest IoU and mean F₁ score, which may be caused by the incomplete and inaccurate pixel ground-truth labels. From the randomly selected test results, it can be found that new damage regions are correctly classified even though they are not appropriately labeled before training. Results shows that the proposed U-net segmentation model can achieve good recognition performance for the irregular concrete spalling and rebar corrosion regions. Some recognition misjudgments are actually regarded as accurate, even though they miss the so-called ground-truth pixels because labeling errors are supposed to exist during the pre-labeling process. This phenomenon also indicates that the trained U-net model has gained good recognition ability for concrete spalling and rebar exposure damages under the circumstances of wrong prior labeling information.