Answers are grounded in this paper and cite their sources.
Garbage classification is the first step in waste segregation, recycling, or reuse. MobileNet was used to generate a model that classifies common trash according to the following categories: glass, paper, cardboard, plastic, metal, and other trash. A dataset of 2527 trash images in .jpg extension was used for the training. The model used transfer learning from a model trained on the ImageNet Large Visual Recognition Challenge dataset. The TensorFlow for Poets git repository was cloned as a working directory to retrain the MobileNet model in 500 steps. The resulting baseline model, with a final test accuracy of 87.2% was optimized and quantized. In the Andoid app development, the optimized model (with 89.34% confidence) is preferred over the quantized model (with 1.47% confidence) based on the test using a plastic image. The model app was successfully installed in a Samsung Galaxy S6 Edge+ mobile phone. The installed mobile app successfully identified a cardboard material in an image with a cardboard container. It is recommended to rerun the training using more steps as this may improve the quantized model performance since a quantized model is fit for mobile devices than models with no quantization.
No takes yet. Share an insight, caveat, or question.
Rabano et al. (2018) studied this question.
Synapse has enriched one closely related paper. Consider it for comparative context: