An Evaluation of Various Deep Convolutional Networks for the Development of a Vision System for the Classification of Domestic Solid Street Waste

Artificial intelligence is an emerging technology revolutionizing the modern world and making the life of mankind easier and more efficient. Its application in various fields is growing every moment like tributaries of flooded rivers. To apply Artificial Intelligence in the field of domestic waste collection, the images of 7 various categories of waste such as cardboards and tetra packs, dairy packets, facemasks, footwear, paper cups, plastic bottles, and wrappers taken in the various factual street environments under natural lighting were trained and tested through different pre-trained convolutional neural networks. This work aims at the development of a vision system for the autonomous robot to collect domestic solid waste littered along streets in densely populated areas where the waste collection process is tedious, by evaluating various existing networks. Image categorization and identification technology is a vital part of the vision system. A total of 700 images consisting of 100 images of each category were used for training and testing purposes. Among them 70% of the images were used for training and 30% of the images were used for testing. The pre-trained convolutional neural networks squeezenet, googlenet, inceptionv3, densenet201, mobilenetv2, resnet18, resnet50, resnet101, xception, inceptionresnetv2, shufflenet, nasnetmobile, nasnetlarge, darknet19, darknet53, efficientnetb0, alexnet, vgg16 and vgg19 were used to evaluate the performance of the image categorization and identification. The testing accuracy of Inceptionv3, densenet201, resnet50, resnet101, xception and efficientnetb0 was 90%. The testing accuracy of nasnetlarge and darknet53 was 91% and the highest testing accuracy of 93% was achieved by inceptionresnetv2.