Classification and object detection models were developed for use in automatic waste-sorting systems, support for waste management in construction projects, environmental protection, supporting the health of living organisms, and error and cost reduction. In this article, two independent and traceable datasets, after duplicate and similarity screening, were modeled using EfficientNet-B0 with transfer learning. The models classified waste images into 11 and 8 commonly used classes with accuracies of 97.03% and 91.31%, respectively. Comparison with a custom CNN trained from scratch showed that the EfficientNet-B0-based model achieved faster convergence in the early training epochs, higher validation accuracy, and a lower training time per epoch. After out-of-sample evaluation and the demonstrated importance of image diversity, the two datasets were merged to create the seven-class MWD7 dataset, and the new model achieved an accuracy of 97.02%. Comparison between EfficientNet-B0 and EfficientNet-B2 showed that EfficientNet-B2 improved the evaluation metrics when trained with its standard input size rather than a fixed input size. This result indicates that heavier architectures in this family, when used with their standard input size, can improve model quality. After the limitations of EfficientNet in analyzing images containing multiple waste types became evident, the multi-object MOWD dataset was created, and YOLO26s was trained to simultaneously detect the type and location of each waste object in multi-object images, achieving an mAP50-95 of 0.792. All models were evaluated using unseen images and multiple metrics. The results showed that the proposed methodology improves the accuracy and efficiency of intelligent waste classification and object detection.