摘要
针对深度学习在水声研究领域的应用中面临大数据量要求和现有样本量限制的问题,本文提出了一种多层优选卷积网络模型。通过基于相似度的优选方法选出最佳卷积核,以提取更具代表性的特征。利用探索层特征融合策略,叠加多层卷积输出,获取更全面的特征信息。采用约减策略优化模型,有效缩短运算时间。通过优选、特征融合和注意力机制,有效解决此类问题。实验结果表明,该模型在数据集上取得的最好的标注准确率为高基线模型1.12%;同时运行时间减少了43.5%。因此,该模型的使用提高了水声信号标注的准确率和效率。
The application of deep learning in underwater acoustic research often faces problems such as large data volume requirements and current sample size limitations.Herein,the best convolution kernel is selected using the similarity-based optimization method to extract representative features.Then,by exploring the layer feature fusion strategy,the multilayer convolution output is superimposed to obtain comprehensive feature infor-mation.This study proposes a multilayer optimized convolutional network model that can effectively solve such problems through optimization,feature fusion,and attention mechanisms.Finally,a reduction strategy is used to optimize the model,which effectively shortens the operation time.The experimental results reveal that the best annotation accuracy of the model on the data set is 1.12%of the high baseline model,and the running time is reduced by 43.5%.Therefore,this model improves the accuracy and efficiency of underwater acoustic signal labeling.
作者
王红滨
张帅
何鸣
陈夏可
WANG Hongbin;ZHANG Shuai;HE Ming;CHEN Xiake(College of Computer Science and Technology,Harbin Engineering University,Harbin 150001,China;College of Computer Science and Information Engineering,Heilongjiang University of Science and Technology,Harbin 150022,China)
出处
《哈尔滨工程大学学报》
EI
CAS
CSCD
北大核心
2024年第4期758-763,共6页
Journal of Harbin Engineering University
基金
基础科研项目(JCKY2019604C004).
关键词
水声信号
自动标注
声纹识别
多层优选卷积模型
时间优化
注意力机制
特征融合
underwater acoustic signal
automatic annotation
voiceprint recognition
multilayer optimal convolution model
time optimization
attention mechanism
character merger