A Review and Application of Deep Learning Method for Classification of Concrete Aggregates and Soil

Mechanical sieve analysis is a common method for determining concrete aggregates and soil classification. For facilitation and acceleration of this method, this paper reviews image processing and deep learning methods used in geotechnical and civil engineering applications. Combination of deep learning with image processing can result in a robust, human-independent approach (in terms of experience and recognition power), resulting in faster and accurate results. To better understand the performance of such methods, two convolutional neural network (CNN) architectures (e.g. AlexNet or GoogleNet) were evaluated for their capability in automatic feature extraction and image classification. It was observed that the accuracy of these networks in prediction of aggregate class is dependent on ratio of the number of training samples to the whole dataset size, epoch number and mini batch size. The number of training images between 80-90% of the total dataset was found to be suitable and a minimum of 10 epoch is required to obtain the maximum validation accuracy. Using this model, a validation accuracy of up to 100% was reachable. Furthermore, the model was capable to predict about 85% of new images correctly. The future improvement of this method can be associated to increasing its efficiency in training process by using optimization approaches.