Application of perceptual image coding and the neural network method in predicting the optimum Asphalt binder content of open-graded friction course mixtures

• Published in: Road materials and pavement design 2017-01, Vol.18 (1), p.38-63
• Main Authors: Mejias de Pernia, Yolibeth, Gunaratne, Manjriker
• Format: Article
• Language: English
• Subjects: Asphalt pavements; human vision system; Image coding; image processing; Neural networks; optimum asphalt binder; pavement; perceptual image coding; pie plate visual method; Roads & highways
• ISSN: 1468-0629; 2164-7402
• DOI: 10.1080/14680629.2016.1139499

Florida Department of Transportation (FDOT) designs open-graded friction course (OGFC) mixtures using a pie plate visual draindown method (FM 5-588). In this method, the optimum asphalt binder content (OBC) is determined based on visual assessment of the superficial asphalt binder draindown (SABD) distribution of three OGFC samples placed in pie plates with pre-determined asphalt binder contents (AC). In order to eliminate the human subjectivity involved in the current visual method, an automated method for quantifying the OBC of OGFC mixtures is developed using digital images of the pie plates and concepts of perceptual image coding and neural networks. Phase I involved the FM-5-588-based OBC testing of OGFC mixture designs consisting of a large set of samples prepared from a variety of granitic and oolitic limestone aggregate sources used by FDOT. Then the digital images of the pie plates containing samples of the above mixtures were acquired using an imaging setup customised by FDOT. The correlation between relevant digital imaging parameters and the corresponding AC was investigated initially using conventional regression analysis. Phase II involved the development of a perceptual image model using human perception metrics considered to be used in the OBC estimation. A General Regression Neural Network (GRNN) was used to uncover the nonlinear correlation between the selected parameters of pie plate images, the corresponding ACs and the visually estimated OBC. GRNN was found to be the most viable method to deal with the multi-dimensional nature of the input test data set originating from each individual OGFC sample that contains AC and imaging parameter information from a set of three pie plates. GRNN was trained by a major part of the database completed in Phase I. Finally, the prediction results from an independent part of the above database demonstrated that the GRNN model provides satisfactory estimations of OBC.