Automatic Annotation of Web Images for Domain-Specific Crack Classification

Abstract

Manual visual crack detection and classification for inspection of civil infrastructure is time-consuming and labor-intensive. Many automatic crack detection and classification algorithms have, thus, been developed in the past decade, several of which achieved acceptable performance results for specific applications and using large datasets for training. However, developing training data for automatic crack classification is not an easy task. It requires a large dataset in terms of quantity and variability, as well as well-trained professionals to label the dataset. Hence, there is a need for efficient ways to develop well-labeled datasets that could not only reduce human effort, but also adapt to diverse inspection contexts for improved classification performance. To address this need, this paper proposes a data retrieval and annotation method to automatically retrieve and label crack images from the Web. The dataset can be used as pseudo training data for supervised machine learning-based crack classification algorithms. The proposed method incrementally retrieves and labels crack images. A weak Convolutional Neural Network classifier first learns from a limited set of Web images, and then acts as a machine annotator and further labels a larger size of data. The proposed method was able to retrieve and label a set of images with 95% labeling recall, which shows that the proposed approach is promising.