Multiple-stage multiple-instance deep learning for cost reduction in annotation of cracks

Abstract

It is expected to improve the accuracy of crack detection using deep convolutional neural networks. However, collecting many images and labeling/annotation of cracks are required, which is a crucial issue to overcome for practical application. Especially, accurate labeling/annotation at the pixel level is an essential task, to construct deep convolutional neural network models for semantic segmentation, which is a task to classify each pixel in an image. Additionally, accuracy of labeling/annotation affects the performance of deep convolutional neural networks directly. In this article, applying multiple-instance learning, which is a weakly supervised learning method, is proposed to reduce the cost of annotation of cracks. Experimental results show that applying multiple-instance learning repeatedly using pseudo labels given by the trained models and combining the trained multiple-stage deep neural network models improves the performance of crack detection.