A Deep Convolutional Neural Network for classifying waste containers as full or not full
Bania J. Fonseca, Ali Felermino D. M. A., Saide Manuel Saide · 2019
There is a common understanding that cleanliness is somehow proportional to the economic development of a country. Thus, in order to become clean, a country needs to have an efficient garbage monitoring system. One important component of such a system is garbage collection time because if we delay emptying the bins, the trash ends up to putting public health at risk. This paper is about creating a Deep Convolutional Neural Networks (DCNNs) based model for classifying a waste container as full or not, so that can be later on used by real-time garbage monitoring systems to process images acquired by cameras installed nearby the trash bins or smartphones. To achieve this, we trained and tested different well-known DCNNs architectures, namely, ResNet34, ResNet50, Inception-v4 and DarkNet53. The models were trained and tested using Repeated K-Fold Cross-Validation, running 5-Fold Cross-Validation 6 times. The results have showed that Inception-v4 outperformed the other models, with near-perfect results: PR-AUC =0.994, F1=0.988, Precision =0.989, Recall =0.987 and ACC =0.987. With these results can be said: a high Precision DCNNs based model was built.