A neural network-based methodology for automated distress classification of highway pavement images

One of the most important elements of an effective pavement management system is the collection and interpretation of pavement surface distress data. Current procedures for carrying out this process typically involve on-site visual inspection and condition evaluation by field personnel. This method is a subjective, slow process that is also labor-intensive, tedious and often dangerous.;Recent developments in automation of this process have principally been based on the application of machine vision and conventional image processing techniques. Although these developments have considerably advanced the state-of-the-art of automated pavement distress evaluation, their performance has been limited by the inherent shortcomings of conventional image processing techniques applied to pavement images.;The objective of this research was therefore to develop and demonstrate the feasibility of an alternative methodology that is based on integration of conventional image processing techniques and artificial neural network models. The research focused on the application of neural network models as pattern classifiers for image interpretation and classification, resulting in the development of neural network-based approaches for automatic thresholding of the images, and for detection and classification of the distresses in each image. Two neural network models were investigated, namely, the multi-layer feed-forward network (MLF) and the 2-stage piecewise linear neural classifier (PLNC). About 250 of the asphalt concrete pavement images acquired by the firm PASCO USA INC. for the US Strategic Highway Research Program (SHRP) were used in this research.;The results obtained have shown that the MLF was able to detect and correctly classify about 98% of the images with transverse and longitudinal cracking, and 86% of those with alligator and block cracking. Slightly less impressive results were obtained with the PLNC, although it did perform as well as the MLF in detection of alligator cracking. A method for computation of severity an extent measures has also been presented. These results have clearly demonstrated the potential for application of the neural network-based approach in pavement surface distress evaluation systems, which was the primary objective of this research.