Activation function based MLP network for shape aggregate classification

The assessment of aggregate quality depends on manual grading together with mechanical filtering through traditional methods. Aggregates need to pass through multiple mechanical and physical as well as chemical tests to verify their compliance with established standards. The evaluation procedures performed by hand proveto be inherently inefficient and subjective and take up too much time. A project seeks to create an image processing system which will classify aggregates into various categories. The classification system employs an artificial neural network (ANN) to analyse images for determining aggregate shapes. The study compares the performance of different training algorithms for the ANN. The study compares the performance of Levenberg Marquardt (LM) against Bayesian Regularization (BR) as training algorithms. The results show that BR training outperforms other methods since it provides better mean square error (MSE) values and enhanced regression outcomes. The combination between BR training method and MLP network delivers optimal performance levels regarding regression accuracy and MSE measurement. Through BR training the network obtained an MSE of 1.2042 and a regression of 0.9892 which confirms its successful ability to classify aggregates throught image analysis. Through this alternative method researchers gain an efficient and objective solution to replace traditional manual classification approaches.