【作者】 战歌；

【导师】 王兴梅； 康敏旸；

【作者基本信息】 哈尔滨工程大学 ， 计算机技术（专业学位）， 2023， 硕士

【摘要】 随着我国海洋经济的发展和国家安全战略的不断升级,对水下目标识别技术的研究愈加活跃。由于水下环境的高度复杂性和水下目标多样性,水下目标识别技术受到极大的挑战,其识别精度仍不够理想,水下数据集的获取和标注也受到种种限制,因此,开展高精度的水下目标识别技术研究是符合我国战略发展需求的重要课题,在军事和民用领域都具有重要的理论价值和实际意义。自监督特征学习方法包括生成式和对比式,生成式自监督特征学习方法只专注于像素空间特征,而对比式自监督特征学习方法能关注更抽象的潜在语义信息。因此,本文致力于深入研究对比式自监督特征学习方法提升水下目标识别任务性能。首先,提出基于双通道自注意力音频编码器(Dual-channel Self-attention Audio Encoder,DSAE)的自监督特征学习方法。将具有丰富低频信息的梅尔滤波器组声谱图特征和关注高频信号的Gammatone滤波器组声谱图特征统一在对比式自监督特征学习中,使编码器学习到结合不同通道特征优势的高级语义特征,并为了增强网络对信息的选择能力,利用局部自注意力机制的水下目标特征提取模块,关注局部特征更好的提取语义信息。在此基础上,为提高DSAE的自监督特征学习方法在下游任务中的识别精度和鲁棒性,引入数据增强策略和正负样本平衡策略,提出基于动态正样本存储的双通道自注意力音频编码器(Dual-channel Self-attention Audio Encoder with Dynamic positive sample Memory Module,DSAE-DMM)水下目标识别方法。利用时频掩码的数据增强策略增加模型学习难度,提升鲁棒性,进一步提出构建动态正样本存储模块,扩增历史时空的嵌入向量作为正样本,并实现动态更新,使模型更充分地学习正负样本特征,平衡正负样本比例以提升模型的识别精度和泛化能力。通过对水下目标数据集进行实验结果分析识别性能、收敛速度和抗噪性能,验证本文提出的基于动态正样本存储的双通道自注意力音频编码器水下目标识别方法具备良好的识别精度和收敛速度,同时,提出的方法在真实环境噪声及不同强度的人为噪声干扰下均表现出较强的鲁棒性,在水下目标识别任务中具有一定的应用潜力和价值。更多还原

【Abstract】 With the development of China’s marine economy and the continuous upgrading of national security strategy,research on underwater target identification technology has become increasingly active.Due to the highly complex underwater environment and diversity of underwater targets,underwater target identification technology faces great challenges,and its recognition accuracy is still not ideal.The acquisition and annotation of underwater datasets are also restricted by various factors.Therefore,conducting high-precision research on underwater target identification technology is an important topic that meets the strategic development needs of China,and has important theoretical value and practical significance in both military and civilian fields.Self-supervised feature learning methods include generative and contrastive approaches.Generative self-supervised feature learning methods only focus on pixel space features,while contrastive self-supervised feature learning methods can focus on more abstract latent semantic information.Therefore,this paper is dedicated to in-depth research on contrastive selfsupervised feature learning methods to improve the performance of underwater target identification tasks.This paper propose a self-supervised feature learning method based on a dual-channel self-attention audio encoder(DSAE),which unifies Mel filter-bank(FBank)features with rich low-frequency information and Gammatone filter-bank(GBank)features that focus on high-frequency signals in contrastive self-supervised feature learning.This enables the encoder to learn advanced semantic features that combine the advantages of FBank and GBank features.To enhance the network’s ability to select information,we use a local selfattention mechanism in the underwater target feature extraction module to better extract semantic information from local features.On this basis,in order to improve the recognition accuracy and robustness of DSAE selfsupervised feature learning method in downstream tasks,a dual-channel self-attention audio encoder with dynamic positive sample memory module(DSAE-DMM)underwater target identification method is proposed.we introduce data augmentation and positive-negative sample balance strategies.We use a time-frequency enhancement strategy to increase the diversity of data samples and enhance robustness.Furthermore,we propose to construct a dynamic positive sample memory module to expand historical temporal embedding vectors as positive samples,and dynamically update them,allowing the model to better learn the longterm dependencies between data and balance the proportion of positive and negative samples to improve recognition accuracy.Through experimental results analysis of underwater target datasets for recognition performance,convergence speed,and noise resistance,we verify that the DSAE-DMM underwater target identification method proposed in this paper has good recognition accuracy and convergence speed.Additionally,our proposed method shows strong robustness to real environmental noise and different levels of artificial noise interference,indicating potential and value in underwater target identification tasks.更多还原