Classification of solid waste using deep learning techniques

Abstract

newline This research work involves designing an efficient and robust algorithm to segregate the waste for further processing and recycling, using novel Deep Learning (DL) techniques. Wastes can be broadly classified into two types namely biodegradable and non-degradable. The task of automatic waste segregation considers the input image to consist of either a single object to be classified into one of the classes or multiple objects in an image. In the second case, the different objects in the input image are identified and classified. The first part of the present research work namely Segregation of Solid Waste Management based on Resnet Method describes the experimental results of using Residual Network and its variants ResNet18, ResNet34 and ResNet50 to classify wastes into one of the six categories of waste namely paper, cardboard, Plastic, glass, metal and other trash using image classification. The next major contribution of this research is Fusion Technology-based Classifiers to Segregate the Solid Waste , which proposes novel approach of an advanced classification model namely DDR-net, which is an enhancement of the ResNext model boosted with double fusion and regularization. The algorithm involves utilizing the combined advantages of fine-tuned ResNext-101 model and a the Resnext-50 model fully trained from scratch, on the chosen dataset, by a double fusion process with ReLU. However, this mechanism suffers from the dying-ReLU problem. To overcome this, Randomized Leaky ReLU has been adopted in the present research work.