Page 170 - 《应用声学》2023年第6期
P. 170

第 42 卷 第 6 期                                                                       Vol. 42, No. 6
             2023 年 11 月                         Journal of Applied Acoustics                 November, 2023

             ⋄ 研究报告 ⋄



                     结合区域提取和改进卷积神经网络的水下

                                                 小目标检测                 ∗




                                    符书楠     1,2    许 枫    1†    刘 佳    1    逄 岩     1,2


                                              (1 中国科学院声学研究所       北京   100190)
                                                (2 中国科学院大学      北京  100049)
                摘要：针对水下小目标信息量有限而难以提取有效特征导致的检测性能不佳问题，提出了一种结合区域提取
                和融合 Hu 矩特征的改进卷积神经网络水下小目标检测方法。该方法包含区域提取和分类两个步骤。首先以
                马尔可夫随机场分割算法为基础进行区域提取，对潜在目标定位的同时降低伪目标对后续分类的干扰；然后
                提取潜在目标区域的 Hu 矩特征并融入卷积神经网络，形成一种形状特征表征能力更强的改进卷积神经网络
                用于分类。声呐实测数据处理结果表明，该方法可以有效提升对水下小目标的发现概率和正确报警率，与其他
                目标检测方法相比，该方法具有更好的检测性能和泛化性。
                关键词：水下小目标检测；卷积神经网络；Hu 矩特征；马尔可夫随机场分割
                中图法分类号: TB566           文献标识码: A          文章编号: 1000-310X(2023)06-1280-09
                DOI: 10.11684/j.issn.1000-310X.2023.06.021

               Integrating region extraction with improved convolutional neural network for
                                         underwater small object detection


                                    FU Shunan 1,2  XU Feng 1   LIU Jia 1  PANG Yan  1,2

                               (1 Institute of Acoustics, Chinese Academy of Sciences, Beijing 100190, China)
                                   (2 University of Chinese Academy of Sciences, Beijing 100049, China)

                 Abstract: Due to the limited feature information of underwater small objects, it is difficult to extract effective
                 features, resulting in poor detection performance. Aiming at this problem, an underwater small object detection
                 method combining region extraction and improved convolutional neural network fused with Hu moment features
                 is proposed. The method includes two stages of region extraction and classification. Firstly, a region extraction
                 method based on the Markov random field segmentation algorithm is used to locate potential objects and
                 reduce the interference of false objects on subsequent classification. Then, Hu moment features of potential
                 object regions are extracted and fused with the convolutional neural network to form an improved network with
                 stronger characterization ability of shape features for classification. The results of sonar data processing show
                 that the method can effectively elevate the detection probability and correct alarm rate of underwater small
                 objects. Compared with the common object detection methods, the proposed method has superior detection
                 performance and generalization.
                 Keywords: Underwater small object detection; Convolutional neural network; Hu moment features; Markov
                 random field segmentation


             2022-07-31 收稿; 2022-11-28 定稿
             中国科学院青年创新促进会项目 (2020023)
             ∗
             作者简介: 符书楠 (1998– ), 女, 江西抚州人, 硕士研究生, 研究方向: 水声信号处理。
              通信作者 E-mail: xf@mail.ioa.ac.cn
             †
   165   166   167   168   169   170   171   172   173   174   175