A neutral network approach to automated classification of cracks in images of highway pavement using vectorized data

The collection and interpretation of highway pavement distress data is an expensive process. This process, in its traditional sense, required many man hours, as technicians have to survey the highway. This process is also quite dangerous and subject to traffic and weather conditions.

Much research has been performed in order to automate the collection and interpretation phases. This research focuses on an alternative methodology for classification of highway pavement distresses. The Learning Vector Quantizer, an artificial neural network architecture, was implemented to classify images of highway pavement according to the longitudinal, transverse, fatigue and block crack types. Data was recorded from vectorized crack segments produced by a segmentation process. The images were recursively broken into smaller blocks, forming a quad-tree type structure. Each block in the image tree was then classified as to the major distress prevalent.

This technique should allow for easier measurement of extent and severity. Also it allows records to be kept on individual pieces of highway for comparison with future monitoring.

The overall results of the classification were satisfactory, with longitudinal and transverse classification outperforming block and fatigue. This was to be expected as block and fatigue crack types are harder to differentiate. With additional research, this approach should prove to be very beneficial to the pavement management field.