Synthesis of neural network architecture for ship pattern recognition based on pre-training technology

The relevance of the article is due to the information and communication support of navigation by monitoring river vessels using video surveillance cameras. The main goal is to recognize ships in images, for which the application of neural networks has potential. The aim of the paper is to study the performance indicators of vessel recognition by means of available pre-trained networks after their additional training for the assigned tasks and to select the most efficient network. The research considers various pre-trained neural networks. The input data for the networks are ship images. The training sample was collected manually and includes two independent DataSets with images of river vessels and many other objects apart from ships. The networks were built and further trained with the aid of Keras and TensorFlow machine learning libraries. The employment of pre-trained convolutional artificial neural networks for pattern recognition problems and the advantages of utilizing such networks over synthesizing a neural network from scratch are presented. The architecture of efficient pre-trained VGG16 neural network is described in detail. An experiment was conducted in additional training of available pre-trained convolutional neural networks for the assigned task. The efficiency of various pre-trained neural networks was evaluated in terms of the percentage of correct pattern recognition cases on the test set. The most efficient neural network for ship pattern recognition tasks has been selected. NASNetMobile and NASNetLarge networks have shown the maximum accuracy. However, the minimum image size that these networks can work with is larger than for other available networks and the great number of parameters in the convolutional layers of these networks causes a significant increase in retraining and operation time than for other available networks. Concurrently, VGG16 neural network with a small number of parameters and a short time for additional training has proven to be highly efficient which is why it is recommended for the purposes of ship pattern recognition.

1. Gol'cova I.A., Gulamov A.A. Informacionnoe obespechenie uchastka zheleznoj do-rogi Izvestija Jugo-Zapadnogo gosudarstvennogo universiteta. Serija: Upravlenie, vychislitel'naja tehnika, informatika. Medicinskoe priborostroenie. 2017;7(2):6–11. Available at: https://swsu.ru/izvestiya/seriesivt/archiv/2_2017.pdf. (In Russ.)

2. Maklakov E.S., Gulamov A.A. Uzel sbora informacii dispetcherskogo centra. Izvestija Jugo-Zapadnogo gosudarstvennogo universiteta. 2018;22(6):136–142. Available at: https://doi.org/10.21869/2223-1560-2018-22-6-136-142. (In Russ.)

3. Maklakov E.S., Gulamov A.A. Optimizacija «poslednih mil'» do udalennyh uzlov dostupa putem primenenija tehnologii LCAS Modelirovanie, optimizacija i informacionnye tehnologii = Modeling, Optimization and Information Technology. 2019;7(3). Available at: https://moitvivt.ru/ru/journal/pdf?id=635. DOI: 10.26102/2310-6018/2019.26.3.039. (In Russ.)

4. Gulamov A.A., Konarev D.I., Cintez arhitektury nejronnoj seti dlja raspoznavanija obrazov morskih sudov. Izvestija Jugo-Zapadnogo gosudarstvennogo universiteta. 2020;24(1):130–143. Available at: https://doi.org/10.21869/2223-1560-2020-24-1-130-143. (In Russ.)

5. Bastiaan Sjardin, Luca Massaron, Alberto Boschetti. Large Scale Machine Learning with Python, Packt Publishing; 2016. 420 p.

6. Adrian Rosebrock. Deep Learning for Computer Vision with Python. PyImageSearch; 2017. 330 p.

7. Cuong Dao-Duc, Hua Xiaohui, Olivier Morère. Maritime Vessel Images Classification Using Deep Convolutional Neural Networks. SoICT. 2015:276–281. Available at: https://doi.org/10.1145/2833258.2833266.

8. Leclerc M., Tharmarasa R., Florea M.C., Boury-Brisset A.C., Kirubarajan T., Duclos-Hindie N., Ship classification using deep learning techniques for maritime target tracking, 21 st International Conference on Information Fusion FUSION. 2018:737–744.

9. Tom Hope, Yehezkel S. Resheff. Itay Lieder Learning TensorFlow: A Guide to Building Deep Learning Systems. O'Reilly Media; 1 edition; 2017. 242 pp.

10. Andreas Mjuller. Vvedenie v mashinnoe obuchenie s pomoshh'ju Python. Rukovodstvo dlja specialistov po rabote s dannymi. Vil'jams; 2017. 480 s. (In Russ.)

11. Sebast'jan Rashka. Python i mashinnoe obuchenie. DMK-Press; 2017. 418 s. (In Russ.)

12. Martín Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, Manjunath Kudlur, Josh Levenberg, Rajat Monga, Sherry Moore, Derek G. Murray, Benoit Steiner, Paul Tucker, Vijay Vasudevan, Pete Warden, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng, Google Brain Tensor flow: A system for large-scale machine learning. Operating Systems Design and Implementation: Proc. 12th Symposium, Savannah, GA, USA, 2016: 265–283.

13. Antonio Dzhulli, Sudzhit Pal. Biblioteka Keras instrument glubokogo obuche-nija. Realizacija nejronnyh setej s pomoshh'ju bibliotek Theano i Tensor Flow. DMK-Press; 2017. 296 p. (In Russ.)

14. Fransua Sholle. Glubokoe obuchenie na Python. Piter; 2018. 400 p. (In Russ.)

15. Aurélien Géron. Hands-On Machine Learning with Scikit-Learn and TensorFlow: Concepts, Tools, and Techniques to Build Intelligent Systems. O'Reilly Media; 2017. 574 p.

16. Ian Goodfellow. Deep Learning (Adaptive Computation and Machine Learning series). The MIT Press; 2016. 800 p.

17. Tariq Rashid. Make Your Own Neural Network. CreateSpace Independent Publishing Platform; 2016. 222 p.

18. Schmidhuber J. Deep learning in neural networks: An overview. Neural Networks. 2015;(61):85–117.

19. Josh Patterson, Adam Gibson. Deep Learning: A Practitioner's Approach.; 2017. 532 p.

20. Sajmon Hajkin. Nejronnye seti. Vil'jams; 2018. 1104 p. (In Russ.)

21. Kaggle: Your Machine Learning and Data Science Community Game of Deep Learning: Ship datasets. 2019. Available at: https://www.kaggle.com/arpitjain007/game-of-deep-learning-ship-datasets (accessed on: 10.03.2021).

22. COCO: Common Objects in Context 2017 Val images. 2017. Available at: http://images.cocodataset.org/zips/val2017.zip (accessed on: 10.03.2021).

23. Michael Taylor. The Math of Neural Networks. Amazon Digital Services LLC – Kdp Print Us; 2017. 168 p.