Lightweight Deep-Learning Based Metal/Plastic Trash Detection with Fused Features

Recently, several approaches have been proposed for the proper management of waste to achieve environmental sustainability and resource conservation. Proper garbage disposal reduces pollution and prevents the contamination of water and soil. This study proposes a deep-learning tool (DLT) for the automatic detection of metal/plastic waste to enhance the recycling process. This work created the DLT by utilizing a Lightweight Deep Learning (LWDL) model to establish a more efficient and less intricate method. The stages in this tool encompass: the collection of trash images and their resizing to 224x224x3 pixels, feature extraction utilizing the selected LWDL model, performance evaluation through binary classification with three-fold cross-validation, and the identification of the top two LWDL models to produce fused deep features (FDF) to enhance classification accuracy. The experimental method employs the SoftMax utilizing individual LWDL features and FDF, and the findings affirm that the proposed tool yields superior metal/plastic waste classification results compared to the individual features. In the future, this technique may be utilized to categorize authentic waste photographs captured with a digital camera.