基于多尺度特征融合的水下细粒度目标识别方法  

Research on underwater fine-grained target recognition method based on multiscale feature fusion network model

作　　者：周欢 张陈卓 吴沛然 王兴梅[1,2] ZHOU Huan;ZHANG Chenzhuo;WU Peiran;WANG Xingmei(College of Computer Science and Technology,Harbin Engineering University,Harbin 150001,China;National Key Laboratory of Underwater Acoustic Technology,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学计算机科学与技术学院,黑龙江哈尔滨150001 [2]哈尔滨工程大学水声技术全国重点实验室,黑龙江哈尔滨150001

出　　处：《应用科技》2024年第3期65-71,134,共8页Applied Science and Technology

基　　金：重点实验室开放基金项目(KY10600220048).

摘　　要：复杂水下环境和光照条件给水下细粒度目标识别带来严峻挑战。本文在光学传感器数据和红外传感器数据的基础上,采用多模态信息融合技术提出基于多尺度特征融合的水下目标识别方法。首先,在3个尺度下分别融合水下目标多模态特征,为了减小水下多模态特征的信息丢失,对低级多模态特征进行融合以获取目标边缘细节信息,考虑水下复杂环境导致目标间的较大差异,融合中级多模态特征以获取提高特征表达能力和准确性,为了利用丰富的语义信息,引入协同注意力机制对高级多模态特征的像素级相关性进行建模,提取细粒度的水下目标类别信息。其次,在解码器中将不同尺度下融合后的特征表示逐步上采样以提高特征的空间信息,并与同尺度特征表示进行相加融合,进一步增强特征表达的准确性和稳定性。最终,将特征表示与标签特征图进行比较,获得像素级的分类结果,完成水下细粒度目标识别任务。通过单模态和多模态水下数据的实验结果对比分析,验证本文提出的方法在水下目标识别任务上取得了更优的识别性能。The complex underwater environment and lighting conditions pose serious challenges to the recognition of fine-grained underwater targets.On the basis of optical sensor data and infrared sensor data,this article proposes an underwater target recognition method based on multi-scale feature fusion using multimodal information fusion technology.Firstly,the multimodal features of underwater targets are fused at three scales.In order to reduce the information loss of underwater multimodal features,low-level multi-modal features are fused to obtain target edge details.Considering the significant differences between targets caused by the complex underwater environment,intermediate multimodal features are fused to improve feature expression ability and accuracy.In order to utilize rich semantic information,a collaborative attention mechanism is introduced to model the pixel level correlation of high-level features and extract fine-grained underwater target category information.Secondly,in the decoder,the fused feature representations at different scales are gradually upsampled,and fused to further enhance the accuracy and stability of feature expression.Finally,the feature representation is compared with the label feature to obtain pixel level classification results.By comparing and analyzing the experimental results of single-mode and muitimodal underwater data,it is verified that the proposed method has achieved better recognition performance in underwater target recognition tasks.

关 键 词：水下图像 目标识别 光学图像 红外图像 多模态 多尺度 特征融合 像素级 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]