文献类型：期刊文献

英文题名：An improved lightweight YOLOv7-tiny-based network for beach litter detection

作者：Zhu, Daoheng[1]; Li, Xiaoyu[1]

机构：[1] College of Electronic and Information Engineering, Guangdong Ocean University, Guangdong, Zhanjiang, China

年份：2026

起止页码：210

外文期刊名：BDAIA 2025 - 2025 2nd International Conference on Big Data Analytics and Artificial Intelligence Application

收录：EI(收录号：20261020204187)

语种：英文

外文关键词：Beaches - Ecosystems - Health risks - Object recognition

外文摘要：Monitoring and cleaning beach litter is a challenging task and time-consuming since it poses major risks to human health and marine ecosystems. Traditional manual survey techniques are slow, inefficient, and have a limited detection range. To solve these issues, we propose an enhanced lightweight model for beach litter identification and categorization based on YOLOv7-tiny. First, we incorporate the NAM attention mechanism into the original ESNet design to improve detection accuracy. This makes it possible for the model to concentrate more effectively on the channel and spatial characteristics of beach debris. Second, we replace the ELAN module in the PAFPN structure with the GhostNetV2 module, which uses fewer model parameters to capture more litter features. Experiments on our custom dataset indicate that, compared to YOLOv7-tiny, the improved model decreases parameters by 40.5%, raises mean average precision (mAP) by 3.7%, and increases frames per second (FPS) by 35%. Tests on public datasets indicate that the improved model achieves a 2.8% higher mAP and a 33% faster FPS than YOLOv7-tiny, confirming its superior accuracy and efficiency. ? 2025 Copyright held by the owner/author(s).