Deep Convolutional Neural Networks with transfer learning for computer vision-based data-driven pavement distress detection. (30th December 2017)

• Record Type: Journal Article
• Authors: Gopalakrishnan, Kasthurirangan; Khaitan, Siddhartha K.; Choudhary, Alok; Agrawal, Ankit
• Is Part Of: Construction & building materials. Volume 157(2017)
• Page Start: 322
• Page End: 330
• Publication Date: 2017-12-30
• Subjects: Pavement cracking -- Digital Image -- Deep learning -- Transfer learning -- Random Forest -- Convolutional Neural Networks; Building materials -- Periodicals; 624.18
• DOI: 10.1016/j.conbuildmat.2017.09.110
• Languages: English
• ISSNs: 0950-0618

Abstract

Highlights: Pre-trained deep Convolutional Neural Networks (DCNN) were used for crack detection. Pavement images sampled from the FHWA/LTPP database were used as datasets. The truncated VGG-16 DCNN was used as a deep feature generator for road images. Various machine learning classifiers were trained using the semantic image vectors. A neural network classifier trained on deep transfer learning vectors gave the best results. Abstract: Automated pavement distress detection and classification has remained one of the high-priority research areas for transportation agencies. In this paper, we employed a Deep Convolutional Neural Network (DCNN) trained on the 'big data' ImageNet database, which contains millions of images, and transfer that learning to automatically detect cracks in Hot-Mix Asphalt (HMA) and Portland Cement Concrete (PCC) surfaced pavement images that also include a variety of non-crack anomalies and defects. Apart from the common sources of false positives encountered in vision based automated pavement crack detection, a significantly higher order of complexity was introduced in this study by trying to train a classifier on combined HMA-surfaced and PCC-surfaced images that have different surface characteristics. A single-layer neural network classifier (with 'adam' optimizer) trained on ImageNet pre-trained VGG-16 DCNN features yielded the best performance.