Evaluation of Performance of Different Optimizers of Convolutional Neural Network in the Classification of Images of Urban Domestic Solid Waste

The inevitable urbanization in the modern world with its exponential population generates highly voluminous domestic waste. Disposal of this waste has become a challenging task. The process of solid waste management is a very tedious one and requires large manpower and poses serious health hazards. Hence, the implementation of automation in the process of collection and segregation of domestic solid waste has become mandatory. In our work, we have tried to identify and classify domestic urban solid waste in their real background using convolutional neural network (CNN), a genre of deep learning. 1892 photos of commonly littered street wastes were taken with their real background in the presence of natural sunlight. The photos of the wastes were grouped into 22 classes and labeled accordingly. These images were trained through transfer learning in the various pre-trained neural networks such as AlexNet, ResNet-18, Places365-GoogLeNet, SqueezeNet, GoogLeNet, ResNet-50, ShuffleNet, MobileNet-v2, NasNet-Mobile, Inception-v3, and ResNet-101. The performance of the different optimizers sgdm, adam, and rmsprop was evaluated in each of these networks for the different initial learn rates. It was found that the overall performance of the optimizers was similar, where 0.001 was the initial learn rate achieved maximum validation accuracy in most of the convolutional neural network pre-trained models. Among all the networks MobileNet-v2 achieved maximum validation accuracy and was able to predict and classify a maximum of 17 classes of waste. Footwear and wrappers were easily identified by most of the neural networks. Cigarette butts, dry flowers, fabric waste, vegetable waste, and wooden waste were never able to be classified by any of the chosen networks.