Semi-supervised learning based on convolutional neural network and uncertainty filter for facade defects classification

Developing a classifier to identify the defects from facade images using deep learning requires abundant labeled images. However, it is time-consuming and uneconomical to label the collected images. Hence, it is desired to train an accurate classifier with only a small amount of labeled data. Therefore, this study proposes a semi-supervised learning algorithm that uses only a small amount of labeled data for training, but still achieves high classification accuracy. In addition, based on the mean teacher algorithm, this study develops a novel uncertainty filter to select reliable unlabeled data for initial training epochs to further improve the classification accuracy. Validation experiments demonstrate that the proposed method can improve the model accuracy from 79.26% to 84.36% compared to the traditional supervised learning algorithm with 10% of labeled data in a dataset. From another perspective, compared to supervised learning algorithm, the proposed technique can help reduce the time and cost for preparing the labeled data.

Automated summary

The study introduces a semi-supervised learning algorithm that trains an accurate classifier using only a small amount of labeled data, and improves classification accuracy by selecting reliable unlabeled data with an uncertainty filter. Validation experiments show that the proposed method significantly enhances model accuracy compared to traditional supervised learning algorithms, while also reducing the time and cost for preparing labeled data.