Features of convolutional neural networks

Vladimir Skripachev, Mikhail Guida, Nikolay Guida, Alexander Zhukov


The article discusses in detail the use of collapsible neural networks (CNNs) for use in solving problems of detecting objects in aerospace images. The structure of collapsing neural networks is revealed, taking into account the specifics of the application. Bearing in mind the creative essence of the problem of object recognition in aerospace images, the high variability of the solution of the task is illustrated. Approaches to the construction of collapsing neural networks are considered, taking into account the key factors of the effectiveness of solving the problem of detecting objects in aerospace images.

Special attention is paid to the internal structure of the construction of collapsing neural networks. The execution of cross-correlation (convolution) operations in networks is described. The mechanism of the shift (step) of the convolution is clearly shown. The application of two-dimensional filters for single-channel and three-channel images is described. The application of image augmentation modes during convolution is analyzed: without addition ("valid" mode), with addition ("same" and "full" modes). The most commonly used filters in CNNs for determining characteristic lines (horizontal, vertical, inclined), filters for improving boundaries (edges, edges), filters smoothing extreme data values are considered. A general conclusion is made about the use of convolutional neural networks and their application strategy, indicating the most acceptable CNNs design for detecting objects in aerospace images.

Full Text:

PDF (Russian)


SATYAJIT MAITRA . ML Strategy for Machine Learning Projects. https://medium.com/@ssatyajitmaitra/ml-strategy-for-machine-learning-projects-7b674e3bbb9, Jul 8, 2019.

Naveen Mathew Nathan S. CNN vs fully-connected network for image processing, Towards Data Science https://towardsdatascience.com/cnn-vs-fully-connected-network-for-image-processing-8c5b95d4e42f •Sep 17, 2019.

Zhang, A., Tay, Y., Zhang, S., Chan, A., Luu, A. T., Hui, S. C., & Fu, J. (2021). Beyond fully-connected layers with quaternions: parameterization of hypercomplex multiplications with 1/n parameters. International Conference on Learning Representations.

V. O. Skripachev, M. V. Guyda, N. V. Guyda, and A. O. Zhukov. Issledovanie svertochnykh neyronnykh setey dlya obnaruzheniya ob"ektov na aerokosmicheskikh snimkakh. International Journal of Open Information Technologies, 10(7):54–64, 2022.

Convolutional Neural Networks cheatsheet https://stanford.edu/~shervine/teaching/cs-230/cheatsheet-convolutional-neural-networks.

Inductive bias site:wiki5.ru. https://livepcwiki.ru/wiki/Inductive_bias.

Aston Zhang, Zack C. Lipton, Mu Li, Alex J. Smola, Brent Werness, Rachel Hu, Shuai Zhang, Yi Tay, Anirudh Dagar, Yuan Tang. Dive into Deep Learning. https://d2l.ai/index.html.

Backpropagation In Convolutional Neural Networks Jefkine, 5 September 2016. ttps://www.jefkine.com/general/2016/09/05/backpropagation-in-convolutional-neural-networks/


  • There are currently no refbacks.

Abava  Кибербезопасность MoNeTec 2024

ISSN: 2307-8162