After nearly 8 months of sample collection, annotation, and model training, we have completed the first testing task of the fine classification and recognition model V1.0 for urban waste. The model covers 18 subcategories under three major categories: recyclable waste, hazardous waste, and other waste. The proportion of correct classification is 91.8%. According to incomplete statistics, this portion (18 subcategories) of waste accounts for 96% of the total dry waste (excluding construction waste).

The image recognition model based on deep convolutional networks can extract different features of objects (transparency, luster, morphology, etc.), and can perceive subtle differences between different categories, laying a good foundation for further fine classification of subcategories in the later stage.

Deep learning models can be used in the later stages, constantly trial and error, accumulate "experience", and continuously update the categories and classification accuracy that can be recognized by the iteration itself.