Using Class Activation Maps on Deep Neural Networks to Localise Waste Classifications

Abstract

A serious global waste crisis is currently in effect which originates from our lack of sense of duty. This can be resolved by automating the separation process using AI empowered by weakly supervised learning. A prototype system was created by using pre-trained CNN models in CV such as VGG, ResNet, MobileNet and DenseNet. The prototype showed promising results by having the best algorithms obtain an F1-score of over 80% on 2 datasets known as TrashNet and MINC. Some algorithms were also quite efficient, reaching over 10FPS while maintaining less than 10Mb. The localisation accuracy generated from the CAMs of the best models has shown to be around 83% on TrashNet and around 69% on MINC. These results show that not only is it possible through AI to accurately and efficiently classify waste through datasets, but it can also be used to integrate accurate localisation via weak supervision for easier data annotation.