Enhanced RecycleNet for Efficient Waste Classification

Segregation of recyclable waste items is one of the crucial aspects of smart cities and their industrial applications. CNN-based machine learning models are widely used to predict and classify image datasets. Traditional deep learning models are fast in training the image dataset, but the classification accuracy is usually too low. Different densely connected CNN architectures are widely used to improve the accuracy in the image waste classification. Despite the remarkable accuracy in such densely connected models, these models often suffer from high computational complexity during the training phase. To overcome this computational complexity, DenseNet121 has been developed, which reduces the training time due to its unique dense block architecture. RecycleNet is a modification of DenseNet121 where the skip connections in the dense block architecture are changed to reduce the computational complexity. In this paper, we propose a unique model called Enhanced RecycleNet, where the skip connections between the dense block architecture are reduced to one-third than in the DenseNet121 model. This unique architecture has improved the model's performance by 46.3% and decreased the trainable parameters from 7 million to about 2.4 million.