Residual CNN + Bi-LSTM model to analyze GPR B scan images

Özkaya U., ÖZTÜRK Ş., Melgani F., Seyfi L.

Automation in Construction, vol.123, 2021 (SCI-Expanded) identifier

  • Publication Type: Article / Article
  • Volume: 123
  • Publication Date: 2021
  • Doi Number: 10.1016/j.autcon.2020.103525
  • Journal Name: Automation in Construction
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, PASCAL, Aerospace Database, Communication Abstracts, ICONDA Bibliographic, INSPEC, Metadex, Civil Engineering Abstracts
  • Keywords: Bi-LSTM, CNN, GPR, Residual connections
  • Ankara Haci Bayram Veli University Affiliated: Yes

Abstract

In this study, the residual Convolutional Neural Network (CNN) with the Bidirectional Long Short Time Memory (Bi-LSTM) model has proposed for the analysis of Ground Penetrating Radar B scan (GPR B Scan) images. GPR characteristics, scanning frequency, and soil type make it very difficult to analyze GPR B Scan images. Also, noise and clutter in the image make this problem more challenging. The proposed method shows high performance in determining the scanning frequency of GPR B Scan images, type of GPR device, and the type of soil. In particular, residual structures and types of Bi-LSTMs connection within the proposed method led to increasing the performance. The metric performance of the proposed method is higher compared to other transfer learning based CNN structures.