<![CDATA[Accurate waste classification plays a vital role in supporting effective waste management and promoting environmental sustainability, especially amid the continuing increase in global waste generation. This study investigates how the presence and removal of image backgrounds influence the performance of deep learning models in automated waste classification. Two Convolutional Neural Network architectures, namely MobileNetV2 and DenseNet169, were evaluated using a dataset comprising 5,054 images across six waste categories: cardboard, glass, metal, paper, plastic, and trash. Each architecture was trained and tested on two dataset variants: original images with backgrounds and images with the backgrounds removed. Model performance was assessed using accuracy, precision, recall, F1-score, and ROC AUC. The results indicate that DenseNet169 consistently outperformed MobileNetV2 across all evaluation metrics. The highest accuracy, reaching 88.33%, was achieved by DenseNet169 when trained on images retaining their original backgrounds. This suggests that background information may provide meaningful contextual features that enhance classification performance. Conversely, removing backgrounds can limit the visual information available to the model and potentially reduce predictive effectiveness. These findings emphasize the importance of carefully considering background characteristics during dataset preparation and model training. Moreover, the study demonstrates that selecting an appropriate model architecture in relation to dataset properties is essential for optimizing classification outcomes. Overall, this research offers practical insights for improving dataset design and model selection in future automated waste classification systems, while contributing to the advancement of scalable and intelligent deep learning-based waste management solutions.]]>
