Due to the growing population and economy, the amount of waste generated each year in developing countries is tremendously increasing. The current waste management system in Ethiopia relies on labor-intensive methods that not only hurt the well-being of the workers but also hinder the overall effectiveness of waste management. To address these challenges, there is an urgent need to implement an automated waste classification system. Leveraging deep learning algorithms, which have demonstrated remarkable success in image recognition tasks, holds great potential for simplifying waste management processes. However, the application of these algorithms in waste image classification remains underexplored and has drawbacks in terms of image dataset diversity & optimization for some waste classification tasks. Additionally, the majority of existing deep learning models have not been adequately tested for real-time predictions using cameras, which is a crucial aspect for the practical implementation. This study aims to bridge these gaps by performing a comparative performance evaluation of different deep learning models in waste image classification and showing their feasibility in real-time using a camera and Raspberry Pi processor for predictions. We have collected diversified images for recyclable wastes (cardboard, glass, metal, paper, and plastic) using different methods and we use deep learning models (ResNet50, Inception V3, and VGG16) as a base model for fine-tuning. By searching for the optimal combination of hyperparameters with a hyperband tuner from Keras, we achieved a validation accuracy of 95.18% with the InceptionV3 model. Additionally, we successfully implemented our best-performing model (IncepitonV3) in real-time. Keywords: Waste management, Waste Image Classification, Optimization, Deep-learning, Fine-tuning, Pre-trained, ResNet50, InceptionV3, VGG16, Hyperband tuner