A lightweight model for plastic classification based on data augmentation

With the popularization of environmental awareness, there is increasing attention in the environmental protection field towards the recycling of plastic waste from household appliances factories. Against this backdrop, spectroscopic technology, particularly the combination of spectroscopic data and deep learning-based classification techniques, has gradually emerged as a key solution to the challenge of plastic classification. However, there has been limited research delving into the lightweight deployment of plastic classification models, which is one of the crucial directions for future investigation. This paper proposes an innovative feature extraction approach that preserves the overall spectral characteristics while locally eliminating fine features caused by noise. This improvement significantly enhances the algorithm's accuracy, resulting in a 2% increase in precision. To augment the sample size of the dataset, we design a plastic spectroscopy generation model (PSGM) for generating synthetic data. By augmenting the dataset with generated spectroscopic data, a further 3% enhancement in the final algorithm's accuracy is achieved. Furthermore, a lightweight plastic classification model (LPCM) is proposed in this paper, occupying only 0.77M of space while maintaining 98% accuracy and a detection speed of 0.003 s. This model not only meets the actual needs of waste electrical appliance recycling in factories but also has the potential to be applied on embedded controllers, demonstrating broad application prospects.