Comparison of Re-trained CNN Models from Pytorch , Keras, and Tensorflow Frameworks for Image Waste Classification

Date of Award


Document Type


Degree Name

Master of Science in Electronics Engineering


Electronics, Computer, and Communications Engineering

First Advisor

Mr. Carlos M. Oppus


A large amount of waste is produced daily in the Philippines which affects the cleanliness and health of the surroundings. To reduce waste, different kinds of recycling practices are implemented throughout various places; however, not all kinds of materials are recyclable. As such, proper segregation is necessary. Automation of waste classification can help reduce manual labor and mistakes done by waste collectors and everyone who contributes to the accumulation of trash. The study trains a Convolutional Neural Network (CNN) model in classifying the wastes into different categories (metal, paper and carton, glass, plastic, and other wastes). The model is created using the three frameworks, TensorFlow, Keras (Theano Backend), and PyTorch. The results are compared in terms of accuracy of the classification along with the amount of time to train it. The research aims to provide insight into the differences and similarities between each framework when creating a waste classification model. The model is created by retraining a base model, InceptionV3, with an additional set of waste images. The number of steps/epochs, and the learning rate are adjusted to observe changes in accuracy and training time. Moreover, training time is compared as well between using CPU and GPU. The study also makes use of TensorFlow and Android to import the waste classifier into a mobile application. TensorFlow is chosen as the framework as it is the most consistent in terms of accuracy and training time. The motivation towards the study is to improve improper waste management, reduce waste cluttering around the streets, and increase the amount of waste recycled to achieve a cleaner environment.

This document is currently not available here.