Deep Learning Based Recyclable Waste Classification Algorithm Research

Along with the continuous development of society,people’s life has been greatly improved compared to the past,and the quality of life is constantly improving,and the consumption activities are becoming more and more diversified.People’s daily life,whether it is the consumption of food or other commodities,will produce the corresponding household waste.Faced with the increasingly prominent problem of the growing scale of domestic waste,the country is actively looking for ways to effectively reduce its impact.Domestic waste is not completely useless;in fact,domestic waste is a very important resource,and most of it can be recycled to obtain the reuse of resources,which is an important constituent basis for sustainable resource development.Therefore,in this paper,we investigate the classification and detection algorithm of recyclable waste based on deep learning,and investigate how to build a specific algorithm that can be practically applied to classify recyclable waste,based on the specific application environment of the algorithm,and the main research contents are as follows:(1)To address the problem of missing data sets for the classification and detection of recyclable waste,this paper creates data sets based on the existing classification standards for recyclable waste.Since there are various types of garbage generated by residents in daily life,more than 7000 images of recyclable garbage are collected in different scenes,and the images are rotated,flipped,cropped and mosaic enhanced to increase data diversity and model robustness.Then,we classify the recyclable waste into six categories:fabric,glass,metal,paper,plastic,and wood,and complete the annotation of the dataset,and select the self-help method to subdivide the recyclable waste classification and detection dataset.Finally,the recyclable garbage classification and detection dataset is formed.(2)Establish an optimization model for recyclable garbage classification and detection based on the YOLOv5s algorithm.Among the three selected models,the YOLOv5s model has a higher average accuracy of 97.4%.Therefore,the YOLOv5s model,which has more advantages in detection accuracy,is used as the benchmark model for recyclable garbage classification detection tasks.Based on this,the model is trained on a dataset and experimental results are analyzed.Based on the experimental results of the YOLOv5s model,it is optimized in three directions.Firstly,a lightweight GhostNet network is used to replace the original backbone network;Secondly,a 4-scale feature fusion detection method was proposed to enhance the model’s ability to detect small targets;Then,a CA attention mechanism was added to enhance the extraction of key feature information.The average accuracy of the optimized YOLOv5s TPC model is the highest among recyclable garbage detection tasks,at 98.4%.The experiment shows that the optimization method proposed in this article is feasible and effective.