基于改进Yolov8n的珊瑚白化图像目标检测

PDF下载HTML

Title:

Target detection of coral bleaching images based on improved Yolov8n

作者:

韩云涛^1,2, 刘宇鹏¹, 胡跃明³, 孙宝鹏¹, 杨佳琪¹

1. 哈尔滨工程大学 智能科学与工程学院, 黑龙江 哈尔滨 150001;
2. 哈尔滨工程大学 三亚南海创新发展基地, 海南 三亚 572000;
3. 中国船舶集团有限公司第七〇三研究所, 黑龙江 哈尔滨 150078

Author(s):

HAN Yuntao^1,2, LIU Yupeng¹, HU Yueming³, SUN Baopeng¹, YANG Jiaqi¹

1. School of Intelligent Science and Engineering, Harbin Engineering University, Harbin 150001, China;
2. Nanhai Innovation and Development Center, Harbin Engineering University, Sanya 572000, China;
3. No.703 Research Institute of CSSC, Harbin 150078, China

关键词:

Yolov8_CSHC; 珊瑚白化检测; 空间金字塔池化网络; 级联分组注意力模块; CIB_C2f模块; 混合注意力变换器; Marjan balance Dataset

Keywords:

Yolov8_CSHC; coral bleaching detection; spatial pyramid pooling network; cascading grouping attention module; CIB_C2f module; hybrid attention Transformer; Marjan balance Dataset

分类号:

TP391.4

DOI:

10.11992/tis.202412019

摘要:

针对海洋环境中珊瑚白化图像特征模糊、背景复杂多变导致的检测精度不足问题，在Yolov8n的基础上，提出了一种基于改进Yolov8n的针对珊瑚白化图像目标检测的Yolov8_CSHC算法。首先，利用防冗余结构紧凑倒置块(compact inverted block, CIB)改进C2f(concatenated feature fusion)模块，减少模型参数量以提高检测速度。其次，在特征融合网络中引入了基于局部注意力增强空间尺度聚合特征的空间金字塔池化网络，可以增强模型对局部细节的感知能力。最后，在特征融合过程中引入级联分组注意力机制，通过将输入特征分割处理，级联输出的方式提高了注意力的多样性和计算效率，使模型可以快速聚焦特征区域。后续引入混合注意力变换器，主要用于单图像超分辨率重建，进一步增强小目标的语义信息和全局感知能力。实验结果表明，在Marjan balance Dataset上，Yolov8_CSHC相较于Yolov8n算法，GFLOPS降低了12%，mAP@0.5-0.95提高了3.6百分点。该算法可以有效地完成海洋珊瑚白化状况的目标检测任务。

Abstract:

To address low detection accuracy in coral bleaching images caused by blurred features and complex backgrounds, we propose an improved Yolov8n-based target detection algorithm for coral bleaching, named Yolov8_CSHC. The C2f(concatenated feature fusion) module improved by using compact inverted block (CIB) optimizes model parameters, enabling faster detection. A spatial pyramid pooling network enhanced with local attention mechanisms is introduced to improve detailed feature perception. During feature fusion, a cascaded group attention mechanism increases attention diversity and computational efficiency, allowing the model to rapidly focus on relevant feature areas. Additionally, a hybrid attention transformer module is applied for single-image super-resolution, enhancing semantic information and global perception of small targets. Experimental results demonstrate that, on Marjan balance Dataset, Yolov8_CSHC reduces GFLOPS by 12% and improves mAP@0.5-0.95 by 3.6 percentage points compared with Yolov8n, effectively detecting coral bleaching in complex marine environments.