Deep learning model for automated detection of efflorescence and its possible treatment in images of brick facades

Abstract

One of the most common pathologies in exposed brick facades is efflorescence, which, although they often have a similar appearance, their effects and way of solving them can range from a one-off cleaning to a repair that involves adding or replacing the material. Therefore, the novel goal of this work is to verify whether it is possible to automate this task of distinguishing what type of intervention each brick needs. To do this, the methodology followed focuses on proposing, training and validating a deep convolutional neural network with the real-time end-to-end method that simultaneously predicts multiple bounding boxes and class probabilities for those boxes. For this, images of 765 building facades will be used, of which 392 were selected, proceeding to label 4704 bricks, resulting in that the model achieved a mAP maximum at epoch 100 with 0.894, which is therefore of interest for the creation of intervention maps.