Deep Clean: A Weakly Supervised Waste Localization System Using Deep Convolutional Neural Network

Abstract

Nowadays, waste dumping on the city streets has become more frequent, especially in developing countries due to the exponential growth of waste generation. This dumping directly affects the city cleanliness, damages resident’s health, and pollutes the surrounding environment, air, and water. The conventional waste dump detection and collection method involve humans who visit the streets and spots and manually confirm if any dump is obtained. This method requires a considerable number of employees and manual operations, which demands a significant amount of energy, time, and money. Additionally, the random appearance of waste dump on streets cannot be controlled through the conventional method. The research proposes the automated waste dump detection and localization method using deep learning to overcome these disadvantages. In this method, the weakly supervised learning approach is implemented using a deep convolutional neural network model. The deep convolutional neural network is trained for two categories: waste and no waste, using a manually constructed dataset and tested for the above categories and localizing waste dump in images. The model performance is evaluated through matrices for classification, and a survey is conducted to assess the accuracy mask generated by the model for waste localization. The precision, recall, F-score, accuracy, and MCC matrices are 0.9708, 0.9848, 0.9778, 0.9776 and 0.9553, respectively. The average score from the survey for generated masks is obtained 3.9. The performance matrices result imply that the model performs outstanding for classification with an accuracy of 97.76 percent and is significantly good for localization with an average score of 3.9. Additionally, the study demonstrates two approaches for the practical application of the implemented model. (i) Citizen oriented approach: It integrates mobile application with the model. (ii) Internet of Things oriented approach: It integrates the existing surveillance system and the model.