Efficient waste management is crucial for sustainable development, especially with the ever-growing volume of municipal and household waste. Existing waste classification systems suffer from several challenges: they achieve a very low recognition accuracy rate, create class imbalance problems, and find it hard to scale-up applicability in real-world conditions. This paper addresses these issues by proposing the InceptionV3 architecture of deep learning. MBWO is utilized to optimize key hyperparameters such as learning rate and dropout rate, batch size on the TrashNet dataset. As a result of this optimization process, it presents outstanding performance metrics such as 97.75% precision, 99.55% specificity, and provides 98.88% precision compared to traditional methods. It categorizes waste materials as paper, plastic, metal, and glass to facilitate efficient recycling processes and reduce the dependence of landfills. This system overcomes the deficiencies of current systems by incorporating both data augmentation and oversampling techniques to handle class imbalance, along with transferring knowledge with CNNs for adaptability in real-world scenarios. Future work will increase the dataset, allow for real-time waste classification, and extend applications to smart city infrastructures while ensuring scalability and applicability in modern waste management challenges.