Recycling for Recycling: RoI Cropping by Recycling a Pre-Trained Attention Mechanism for Accurate Classification of Recyclables

Abstract

Automated classification of recyclable waste is necessary to process a huge amount of recyclables for reuse. This research features recycling a pre-trained attention mechanism for cropping region of interest (RoI) for efficient classification of recyclable waste. We report that an attention mechanism pre-trained with the MNIST dataset, followed by simple morphological operations, successfully provides a bounding box for a recyclable object to be fed into object recognition models such as ResNet50 and EffNetB0. This way, we avoid the cost of annotating large datasets to train state-of-the-art object detection models such as YOLO and R-CNN. Experimental results using the Recyclable Solid Waste Dataset (RSWD) show that our attention-based RoI cropping method is effective enough to separate an object for recognition to achieve accurate classification of recyclables.