The transfer learning with convolutional neural network method of side-scan sonar to identify wreck images

被引：0
|
作者：
Tang Y. [1 ]
Jin S. [1 ]
Bian G. [1 ]
Zhang Y. [1 ]
Li F. [1 ]
机构：
[1] Department of Hydrography and Cartography, Dalian Naval Academy, Dalian
来源：
Cehui Xuebao/Acta Geodaetica et Cartographica Sinica | 2021年 / 50卷 / 02期
基金：
中国国家自然科学基金;
关键词：
Convolutional neural network; Image recognition; Side-scan sonar wreck image; Transfer learning; VGG-16;
D O I：
10.11947/j.AGCS.2021.20200187
中图分类号：
学科分类号：
摘要：
The Side-scan sonar image automatic recognition is an important part of verification for underwater obstacle and wreck search and rescue, in view of the traditional artificial interpretation of side-scan sonar image is inefficient, time consuming and resource consumption and strong subjective uncertainty and excessive reliance on experience. This paper attempts to introduce the method of convolutional neural network, considering that the side-scan sonar shipwreck image belongs to a small sample data set, and an automatic recognition method of side-scan sonar shipwreck image based on transfer learning is proposed.The sample data were expanded by means of normalization and image enhancement, the training set and testing set were divided into 4:1 proportions, and an improved model was designed according to the characteristics of the side-scan sonar wreck data set by referring to the classical VGG-16 model, then, the improved model trained on the ImageNet image data set is used to learn and experiment on the small sample side-scan sonar shipwreck data set using two transfer learning methods: freeze and train and fine-tuning, and compared with new learning. The results show that the accuracy of the three methods for the recognition of side-scan sonar shipwreck images is 93.71%, 84.49% and 90.58%, respectively. The first transfer learning method has the highest accuracy rate, the fastest model convergence speed, and the highest AP value 92.45%, which is 8.06% and 3.06% higher than the second transfer learning and the new learning method, respectively, and has a better effect in improving the model's recognition ability and training efficiency. which verifies the effectiveness and feasibility of this method and has certain practical guiding significance. © 2021, Surveying and Mapping Press. All right reserved.
引用
收藏
页码：260 / 269
页数：9
相关论文
共 41 条
  • [1] LEHARDY P K, MOORE C., Deep ocean search for Malaysia airlines flight 370, Proceedings of 2014 Ocean-St, pp. 1-4, (2014)
  • [2] XIAO Fumin, LIU Yanchun, BAO Jingyang, Et al., Introduction for hydrography, pp. 1-3, (2016)
  • [3] ZHAO Jianhu, LI Juanjuan, LI Meng, Progress and future trend of hydrographic surveying and charting, Journal of Geomatics, 34, 4, pp. 25-27, (2009)
  • [4] CAO Huifen, Development status of deep-sea exploration technology and equipment in China, Ship Supplies and Marketing, 2, pp. 19-22, (2005)
  • [5] WANG Tonghao, PENG Xingguang, PAN Guang, Et al., Development and key technologies of unmanned underwater vehicles, Aerospace Systems Engineering Technology, 1, 4, pp. 52-64, (2017)
  • [6] ZHANG Lihua, LIU Xianpeng, JIA Shuaidong, Et al., A line-surface integrated algorithm for underwater terrain matching, Journal of Geodesy and Geoinformation Science, 2, 4, pp. 10-20, (2009)
  • [7] LIU Chenchen, LI Chen, ZHANG Zhimeng, PCNN-based texture feature extraction of sonar images, Process Automation Instrumentation, 30, 2, pp. 18-20, (2009)
  • [8] LIU Zhuofu, SANG Enfang, Sonar image recognition based on texture, Computer Engineering, 30, 14, pp. 113-115, (2004)
  • [9] LI Haibin, TENG Huizhong, SONG Haiying, Et al., Technology on the extraction of seabed target based on high resolution side-scan sonar, Hydrographic Surveying and Charting, 30, 6, pp. 71-73, (2010)
  • [10] HE Hao, WANG Shuyang, WANG Shicheng, Et al., A road extraction method for remote sensing image based on encoder-decoder network, Journal of Geodesy and Geoinformation Science, 3, 2, pp. 16-25, (2020)
12345