Mixed waste classification based on vision inspection / Hassan Mehmood Khan

Hassan Mehmood , Khan (2022) Mixed waste classification based on vision inspection / Hassan Mehmood Khan. Masters thesis, Universiti Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (171Kb)
	PDF (Thesis M.A) Download (2030Kb)

Abstract

Classification of dry waste garbage is crucial since incorrect labelling of dry waste types may contribute huge loss to waste industry. An automated garbage sorting conveyor system is developed on image analysis of dry waste garbage samples which involves image acquisition, feature extraction and classification. In this study, an Automated Sorting Conveyor (ASC) integrated with Garbage Image Analysis (GIA) System with capabilities to classify and sort multiple types of garbage autonomously i.e., Crumble (Paper/Plastic), Flat (Paper/Plastic), Tin Can, Bottle (Plastic/Glass), Cup (Paper/Plastic), Plastic Box, Paper Box. A total of 640 samples of image data was collected, out of which 320 image data was used for training of machine learning model while the remaining 320 image data was used for testing purposes. Feature selection was also carried out to find the most relevant features with respect to dry garbage of interest. First, 40 features were selected with training accuracy of 79.59%. Then, better accuracy was obtained when redundant features were removed which accounted for 20 features with 81.42%. Finally, 17 features were tested and excellent accuracy of 90.69% was obtained. However, when the features F1 and F2, were removed which left with 15 features, the accuracy was reduced to 81.83%. The best 17 resulting features were used for the next process. Four classification algorithms specifically the Cubic SVM (C.SVM), Quadratic SVM (Q.SVM), Ensemble Bagged Trees (EBT) and k-Nearest Neighbor (kNN) are employed to test the classification accuracy. The Q.SVM achieved the highest training accuracy of 90.69% with 17 features in the application. Q.SVM was used for 320 testing images with the overall testing accuracy of 89.9% and the result was promising for the implementation of an ASC which is eventually crucial to cater mass recycling activities as a replacement for manual sorting.

Actions (For repository staff only : Login required)