Informativeness-guided active learning for deep learning-based facade defects detection

Annotation work is burdensome and challenging for developing a facade defects detector, especially when the raw data set is large but not all useful. To alleviate the problem, this study proposes an informativeness-guided active learning methodology to effectively select informative data to train a robust facade defects detector. A novel data annotation workflow is developed to ensure the high quality of labels. Then, an active learning-based model training strategy is adopted to enable the model to have both the abilities of generalization and discrimination on different defect features. Besides, an innovative informativeness assessment method is proposed by flexibly combining the degree of uncertainty and the degree of representativeness. Through the proposed method, the performance of facade defects detection can be further boosted with the same amount but more informative training data so that the cost-efficiency of human annotation work can be improved.