This dissertation project aims to develop an image-based Municipal Solid Waste (MSW) classification system that can separate recyclable materials from nonrecyclable ones and enable a self-sorting MSW system. In the past literature, deep neural networks have been used extensively in various computer vision tasks, including image classification and material recognition. However, the success of such deep models is largely driven by high- performance computers and large-scale annotated data. While applying deep models over image based MSW classification, a major challenge lies on the fact that testing environments often have significantly different settings (e.g., lighting conditions) from training data. As a result, a pre-trained deep model might not generalize well to testing images. This dissertation explores a comprehensive approach for synthesizing images of MSW objects using pre-made 3D object models and develop approaches to train deep models from these synthetic images. To mitigate the poor generalization issue, images of testing environment are used as the background of the generated images. Experiments with comparisons to alternative methods were reported on a newly collected image benchmark, which includes thousands of images of multiple MSW objects captured in varying lighting settings. Results with ablation analysis suggested that deep visual models trained from synthetic images can largely improve system performance images of unseen environments.