垃圾
计算机科学
帕斯卡(单位)
人工智能
目标检测
特征提取
跳跃式监视
可视化
特征(语言学)
计算机视觉
分类器(UML)
最小边界框
对象(语法)
数据挖掘
模式识别(心理学)
感知
元组
一般化
理论(学习稳定性)
瓶颈
垃圾收集
可靠性(半导体)
深度学习
降维
亮度
作者
Zhenyu Ke,Zihang Luo,Daizhi Liao,Jiayu Li,Zewen Xie,Peiyi Lu,Guanquan Zhu,Cheng Hu,Changhong Liu
标识
DOI:10.1088/1361-6501/ae5de8
摘要
Abstract Floating garbage on water surfaces presents persistent challenges for environmental monitoring and maritime operations, where reliable visual perception is often degraded by specular reflections, surface ripples, and background clutter. Existing one-stage detectors often struggle to maintain stable discriminative representations and accurate localization under such interference, especially when reflection-induced pseudo-structures and blurred boundaries are present. To support robust sensing under realistic conditions, we construct an extended water-surface garbage dataset by enhancing the ORCA dataset in both scene diversity and object categories. From a measurement-oriented perception perspective, we propose HDP-YOLO, a single-stage detection framework built upon YOLOv8s, which serves as a mature and efficiency-balanced baseline for task-oriented enhancement in aquatic scenarios. The model integrates a high-frequency enhanced feature extraction module to strengthen edge and texture perception, together with a dual-directional multi-scale feature fusion strategy that enhances cross-layer semantic consistency. In addition, a composite IoU-based loss function is introduced to achieve stable bounding box regression under ambiguous object boundaries and reflective disturbances. Experimental results demonstrate that HDP-YOLO achieves a 4.1% increase in precision, 3.1% in recall, 3.5% in F 1-score, 3.4% in mAP@0.5, and 2.5% in mAP@0.5:0.95 over YOLOv8s, and consistently outperforms representative two-stage detectors, recent YOLO variants, and other mainstream one-stage models under comparable computational budgets. Furthermore, cross-dataset evaluations on Pascal VOC, URPC2020, and SODA confirm the generalization capability of the proposed method.
科研通智能强力驱动
Strongly Powered by AbleSci AI