A Deep Learning Approach to Vegetation Images Recognition in Buildings: a Hyperparameter Tuning Case Study

Abstract

Deep Learning methods have important applications in digital image processing. However, the literature lacks further studies that propose machine learning models to images classification in civil construction area. For example, the vegetation recognition on facades can be relevant in identifying the degradation and abandonment of buildings. Thus, the objective of this paper is to propose an Convolutional Neural Networks (CNN) approach to vegetation images recognition in buildings. For this, a database with urban images (low altitude) captured by a drone in Zurich (Switzerland) was adopted. In addition, a rigorous hyperparameters tuning methodology for the CNN model is presented. After adjusting the hyperparameters and the final model, the system achieved 90% of accuracy in the test stage. It should also be noted that CNN correctly classified 97.8% of the positive class (with vegetation on the facade) in test images.