A Sunken Litter Detection using Dual Receptive Excitation Module

Abstract

Sunken litter poses a severe ecological challenge, threatening marine life and global ecosystems. Plastic litter is particularly concerning as it could disrupt the food chain, impacting the biodiversity and ecosystem. Over time, without intervention, this issue poses a severe threat to global food security, economic stability in coastal communities, and overall environmental balance. Addressing this problem requires effective monitoring systems for detection. This study enhances the YOLOv10 architecture with a novel Dual Receptive Excitation (DRE) module to improve sunken litter detection. The DRE module uses a dynamic dual-kernel approach to balance spatial and channel-wise processing in Convolutional Neural Networks, adaptively adjusting the receptive field, and capturing critical patterns across scales. Evaluations on the challenging Trash-ICRA19 dataset, sourced from J-EDI, demonstrate the model's robustness under diverse underwater conditions. The proposed system achieves a mean average precision (mAP) of 47.4% and processes 19.60 frames per second, outperforming other studies.