Image Retraining Using TensorFlow Implementation of the Pretrained Inception-v3 Model for Evaluating Gravel Road Dust

Abstract

In gravel roads management systems (GRMS), the need for a holistic approach for detecting the dust amounts on gravel roads has enabled the development of a solution that works based on one of the subdisciplines of artificial intelligence (AI). Recently, machine learning is one of the most widely used algorithms to train data to optimize systems. The advances in machine learning has enabled us to develop a complex application. This paper demonstrates the ability of using one of the most popular machine learning frameworks TensorFlow to build an image classifier. This classifier has the ability to classify the dust amounts on gravel roads into four major levels (None, Low, Medium, and High). This classifier is based on the aspect of optimizing one of the deep neural networks models Inception-v3 model. This model contains a pretrained package used to extract and recognize dust patterns from dust images automatically. In this paper, a data set of 4,000 images of gravel roads were collected. For training, 80% of the data set was used, and 20% was used for testing. Furthermore, a prediction accuracy plot was generated, and it was found that this classifier achieves a prediction accuracy of 72%.