Segmentation by Teacher-Student Adaptive Deep Learning and Its Application to Diagnosis of Deterioration of Power Transmission Tower Images

Abstract

Using the Restricted Boltzmann Machine (RBM) and Deep Belief Network (DBN), a structural adaptive deep learning method was proposed to automatically find the number of hidden neurons and the number of hidden layers considered appropriate for the input data space during learning. Furthermore, a Teacher-Student (T/S) structural adaptive learning method was proposed as an ensemble learning model when there is a mixture of label assignments by multiple annotators, for example, when different teacher signals are assigned to the same input signal. The proposed method has shown higher classification accuracy than conventional methods on various image classification benchmark datasets and has provided an alternative to many time and efforts, such as medical image diagnosis and image recognition in civil engineering and architecture fields. In this paper, we applied this method to the automatic diagnosis of deteriorated areas in the photographs for exterior inspection of power transmission towers using drones. We developed a visualization function based on rectangular detection of abnormal areas and segmentation according to the state of the abnormality. The experimental results showed that rectangle detection was performed with 97.57% accuracy.