融合类别数量自适应深度数据增强和迁移学习的造礁珊瑚识别方法研究  

Integration of category-quantity adaptive deep data augmentation and transfer learning for reef-building coral recognition

作　　者：王岚 魏皓 车亚辰 张翠翠 Wang Lan;Wei Hao;Che Yachen;Zhang Cuicui(School of Marine Science and Technology,Tianjin University,Tianjin 300072,China;Key Laboratory of Ocean Observation Technology of Ministry of National Resources,National Ocean Technology Center,Tianjin 300112,China;Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]天津大学海洋科学与技术学院,天津300072 [2]国家海洋技术中心自然资源部海洋观测技术重点实验室,天津300112 [3]中国科学院计算技术研究所,北京100190

出　　处：《海洋学报》2024年第9期120-130,共11页

基　　金：国家重点研发计划项目(2022YFC3104600);海南省重点研发计划项目(ZDYF2024SHFZ051);国家自然科学基金面上项目(42076007);自然资源部海洋观测技术重点实验室定向基金(klootB06)。

摘　　要：造礁珊瑚识别对于珊瑚礁生态系统的保护与监测具有重要意义。深度学习作为图像识别的前沿技术,在珊瑚识别领域逐渐得到应用。然而,其识别性能仍然面临挑战。其中,数据集中类别间样本数量不平衡和数据多样性欠缺是两个主要问题。前者使得深度学习模型在特征提取过程中更偏向于样本数较多的类,对少数类(尤其是濒危珊瑚)的学习能力不足进而影响其识别准确度。后者因为数据缺乏多样性使得模型无法充分学习各种珊瑚特征,进而限制了特征提取的能力。鉴于此,本文提出了一种融合类别数量自适应深度数据增强和迁移学习的造礁珊瑚类型识别方法。针对第一个问题,本文利用识别结果评价指标F_(1)-score定义的数据生成量化公式对原始深度数据增强方法DeepSMOTE进行改进,提出了类别数量自适应的深度数据增强方法DeepSMOTE-F_(1)。该方法根据每类珊瑚的识别结果自适应地增强其样本数量,确保模型充分学习各类珊瑚特征。针对第二个问题,利用迁移学习强化了模型的提取能力。实验结果表明,在RSMAS、EILAT和EILAT2这3个代表性珊瑚识别数据集上,相较于原始DeepSMOTE,本文提出的DeepSMOTE-F_(1)识别准确率分别提升了2.88%、0.39%和1.54%;与现有的珊瑚智能识别方法相比,准确率分别提升了0.76%、1.40%和1.30%。Recognition of reef-building corals is important for protecting and monitoring coral reef ecosystems.Deep learning,as an advanced technology in image recognition,has been increasingly applied in coral recognition.However,its performance is still challenged by several issues,such as the imbalance of samples among different coral categories within a dataset and the limitation of data diversity.The former makes the deep learning model more likely to extract features from classes with a large number of samples and,therefore,decreases its ability to re-cognize small-sample-size corals,which often refer to endangered ones needing to be protected.The latter further reduces the performance of deep learning in recognizing corals with different appearances and are captured in vari-ant environments.To solve these two problems,this study develops a reef-building coral recognition method by in-tegrating a category-quantity adaptive deep data augmentation algorithm and transfer learning.To address the first problem,a category-quantity adaptive deep data augmentation algorithm named DeepSMOTE-F_(1)is proposed.This algorithm improves the existing DeepSMOTE by introducing a sample-size determination stagey using an F_(1)-score based evaluation metric.It can adaptively augment the number of samples of each category of corals according to its recognition performance so that the deep learning model can fully learn features from each class of corals.For the second problem,transfer learning is used to further enhance the model's ability to extract features.The experi-mental results on three widely used public coral recognition datasets,RSMAS,EILAT,and EILAT2 show that the recognition accuracy of the proposed DeepSMOTE-F_(1)is improved by 2.88%,0.39%,and 1.54%,respectively,com-pared with the traditional DeepSMOTE;and the accuracy of the integrated method is improved by 0.76%,1.40%and 1.30%compared with the existing deep learning methods for coral recognition.

关 键 词：珊瑚识别 深度学习 数据集不平衡 数据增强 迁移学习 

分 类 号：P714.5[天文地球—海洋科学]