Household Waste Segregation Using Intelligent Vision System

Teh, Junjie (2020) Household Waste Segregation Using Intelligent Vision System. Final Year Project, UTAR.

Preview

PDF Download (1686Kb) | Preview

Abstract

Waste segregation is a process to separate the wastes accordingly into their specific categories. Each waste goes into its category at the point of dumping or collection. Waste segregation is an important step to improve the effectiveness of waste management. Normally, this process is done manually by human hand picking in Malaysia. The waste generated in Malaysia is increasing gradually every year, the environment will be polluted if the waste is not managed properly and it will also endanger human’s health. Thus, an intelligent vision system is proposed to improve the efficiency of waste segregation. According to literature review, CNN appeared to be a promising way to develop an intelligent vision system for waste segregation. However, the CNN models take a long time to train and predict. In this project, a study on the types of household waste generated in Malaysia is conducted. The identified wastes are being divided into 6 different classes for this experiment which are glass, metal, cardboard, plastic, paper, and other wastes. Next, several famous CNN architectures such as VGG-19 and Inception V3 are studied and experimented in this project to benchmark with the state-of-the-arts. Besides that, we have proposed a novel method which is the hybrid CNN-ELM model. The hybrid model aimed to improve the efficiency of the system in real-time application. The training and predicted time of the hybrid CNN-ELM model is 720 times faster than the conventional CNN architectures. All the models built in this project are tested by two different publicly available databases which are the Trash-Net dataset and the 0528qsw dataset. The test accuracy of VGG-19 is the best among the others which scored above 90% in both datasets. However, the InceptionV3+ELM model can achieve an accuracy of 90% in the 0528qsw dataset. The proposed hybrid CNN-ELM model has higher computational efficiency compared to the conventional deep learning methods as the time taken for the model to compute is only 5.4s whereas the VGG-19 model takes 2954s.

Actions (login required)