基于层次结构与多模块的海洋生物分类算法  

Marine Organism Classification Algorithm Based on Hierarchies and Multi-module

作　　者：于升正 程远志 YU Sheng-zheng;CHENG Yuan-zhi(School of Information Science and Technology,Qingdao University of Science and Technology,Qingdao 266100,China;School of Information Science and Technology,Harbin Institute of Technology,Harbin 150001,China)

机构地区：[1]青岛科技大学信息科学技术学院,山东青岛266100 [2]哈尔滨工业大学信息科学技术学院,黑龙江哈尔滨150001

出　　处：《计算机技术与发展》2024年第11期36-42,共7页Computer Technology and Development

基　　金：国家自然科学基金(62172249)

摘　　要：传统分类方法在海洋生物图像分类任务上视各类别相互独立,而生物间存在着明确的相互关系,常规方法忽略了其生物学关系。为了使分类网络充分利用数据间的关系,该文提出层次化贝叶斯信息准则(HBIC)探索分层结构,并结合预定义层次结构联合学习,共同辅助神经网络分类。此外,为更高效准确地提取数据全尺寸特征,设计了一种EAConv模块,并引入相对注意力机制,基于多模块与层次结构,进一步建立端到端联合优化的分层学习方法框架(EAHNet)。所有实验基于私有的南麂列岛潮间带大型海洋生物数据集进行,根据层次结构设计的常规卷积神经网络能够将分类准确率提高到86.16%,完整网络能够使准确率达到96.17%,同时能够保证准确率与参数量等网络性能的均衡。结果表明,所提出的多种层次结构辅助、卷积与注意力机制特异性结合的特征提取方法,有效加强了网络对于海洋生物关系信息与特征的捕获能力,从而在整体上取得非常有竞争力的结果。The application of deep learning in marine life classification is crucial for marine ecological research.Traditional methods treat each category independently,while there are clear interrelationships among organisms.In order to fully utilize the relationships between data in classification networks,we propose the Hierarchical Bayesian Information Criterion(HBIC) to explore hierarchical structures and combines predefined hierarchical structures for joint learning to assist neural network classification.Additionally,to extract data features efficiently and accurately across all dimensions,a novel EAConv module is designed,incorporating a relative attention mechanism.Based on multiple modules and hierarchical structures,an end-to-end joint optimization framework for hierarchical learning(EAHNet) is further established.All experiments are based on a proprietary dataset of large marine organisms from the intertidal zone of Nanji Islands.A conventional convolutional neural network designed based on hierarchical structures improves classification accuracy to 86.16%,while the complete network achieves an accuracy of 96.17%,ensuring a balance between accuracy and network performance metrics such as parameter count.The results demonstrate that the proposed methods,which combine various hierarchical structures,convolution,and attention mechanisms,effectively enhance the network's ability to capture information and features related to marine organisms,resulting in highly competitive overall performance.

关 键 词：层次结构 层次化贝叶斯信息准则 联合优化 多模块 海洋生物图像 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]