摘要
随着水下资源的开发,深度学习在水资源探索和开发领域应用越来越广泛。在水下原始图像质量低下的情况下,传统的语义分割技术对水下目标分割边界模糊、定位不准确、漏检和误检的情况经常发生。论文针对上述问题,提出了一种专门应用于水下图像的语义分割方法。首先基于多空间转换对原始水下图像进行增强处理。其次通过密集连接的混合空洞卷积在扩大感受野的同时消除多层空洞卷积带来的“gridding issue”问题,然后设计级联空洞卷积空间金字塔池化模块来整合不同尺度的边界特征,丰富目标细节信息。最后,采用上下文信息聚合机制将浅层网络和深层网络的特征进行融合以提取丰富的上下文信息。实验证明论文提出的方法相比最先进的语义分割方法对水下图像的分割效果更好。
With the development of underwater resources,deep learning is more and more widely used in the field of water resources exploration and development.Under the condition of low quality of the original underwater image,the traditional semantic segmentation technology often makes the underwater target segmentation boundary fuzzy,inaccurate positioning,missed detection and false detection.Aiming at the above problems,this paper proposes a semantic segmentation method for underwater images.Firstly,the original underwater image is enhanced based on multi-space transformation.Secondly,the problem of"gridding issue"caused by multi-layer hole convolution is eliminated by densely connected mixed hole convolution.Then,a cascaded hole convolution space pyramid pooling module is designed to integrate the boundary features of different scales and enrich the detailed information of the target.Finally,the features of the shallow network and the deep network are fused by the context information aggregation mechanism to extract rich context information.Experiments show that the proposed method is better than the most advanced semantic segmentation method in underwater image segmentation.
作者
王金康
何晓晖
邵发明
卢冠林
WANG Jinkang;HE Xiaohui;SHAO Faming;LU Guanlin(Army Engineering University of PLA,Nanjing 210007)
出处
《舰船电子工程》
2023年第9期29-35,共7页
Ship Electronic Engineering
关键词
水下图像
语义分割
编解码机制
上下文信息聚合
underwater image
semantic segmentation
encoder-decoder architecture
context information aggregation
作者简介
王金康,男,硕士研究生,研究方向:深度学习、图像处理;何晓晖,男,博士,教授,研究方向:机械智能化;邵发明,男,博士,副教授,研究方向:软件工程、信号处理;卢冠林,男,硕士研究生,研究方向:图像处理。
