Convolutional Neural Network for Automatic Trash Classification

This thesis is about adapting a previously built mechanism called ATC that in its present state can place trash into five different categories (plastic, paper, aluminium cans, plastic bottles and general waste) by pressing a button for the selected category. Now by implementing a Convolution Neural Network to the mechanism an attempt will be made to make this process automatic and therefore the push buttons irrelevant. Datasets for four categories will be made in the camera chamber of the ATC mechanism duplicating the future environment for analysis. Dataset pre-processing with resizing and image augmentation will be discussed as well as splitting the data to training, validation and test sets. The main features for building, training and testing a Convolution Neural Network are discussed with my decisions and reasoning on selected parameters. Brief history of Neural Networks, Convolution Neural Networks and some successful architecture are reviewed. The top model accuracy and loss function on the training, validation and test data are displayed visually with confusion matrices and graphs along with comparison with another model