Development of Datasets to Detect and Classify the Waste by Using Deep Learning

One of the biggest problems with today's environment is waste pollution. Everyone knows that recycling is essential for the environment and the economy and that recycling companies need to be very efficient. Additional standards and benchmarks to measure the data are utilized in current research on autonomous waste identification, making it almost impossible to compare results. To solve these issues, this paper reviews the current waste detection methods that rely on Deep Learning and conducts a rigorous study of more than ten waste datasets. This paper aims to provide the first reproducible benchmark for litter detection by collecting and summarizing prior research and providing the outcomes of the authors' experiments on the submitted datasets. In addition, two novel benchmark datasets are suggested, to detect and classify waste, amalgamations of the aforementioned open-source datasets annotated consistently across all potential waste types: glass, metal and plastic, non-recyclable, bio, other, paper, and unfamiliar. Lastly, a litter localization and categorization detector with two stages is introduced. The localization of litter is accomplished using EfficientDet-D2, and garbage detection is classified into seven groups using EfficientNet-B2. The classifier is taught using unlabelled photos in a semi-supervised way. On the test dataset, the suggested method attains an average waste detection and classification precision of 75%. Anyone may access the research’ code and notes on the internet.