军事知识服务平台
基于改进分割注意力残差网络的珊瑚种类图像识别方法
Method for Coral Species Image Recognition Using Enhanced ResNeSt
摘要 针对珊瑚图像识别中类间差异小、边缘模糊及数据不平衡导致分类精度低等问题,提出了一种基于分割注意力残差网络(Split-Attention Networks,ResNeSt)的分类方法.该方法通过卷积注意力机制提升特征提取能力,解决了类间差异小和边缘模糊问题;使用焦点损失函数处理数据不平衡,并通过混合数据增强策略防止过拟合,提高了模型鲁棒性.验证结果表明:该方法在南海珊瑚图像数据集上取得了 88.86%的测试精度,在常用的珊瑚分类网络中表现出最优的性能.
Abstract:
Aiming at the problems of low classification accuracy caused by small inter-class differences,blurry spatial edges and data imbalance in coral image classification,this paper proposes a coral classification model based on ResNeSt.To address the problems of small class differences and fuzzy edges among corals,we introduce the convolutional attention mechanism to enhance useful features and suppress redundant features,which improves the feature extraction ability of the model.To address the data imbalance and insufficiency in our dataset,we adopt the Focal Loss function to mitigate the imbalance issue.Additionally,we employ Mixup data augmentation to alleviate overfitting caused by limited data,thereby enhancing model robustness.The model finally achieves 88.86%accuracy on the South China Sea coral image dataset,which outperforms the commonly used coral classification networks.
刊 名 海洋技术学报 ISTIC
年,卷(期): 2026, 45(4)
分类号: TP753
机标分类号: TP391.41 TS805 TP181
在线出版日期: 2026年9月8日
基金项目: ZDYF2023GXJS023:海南省重点研发计划资助项目,