Upgrade to Pro — share decks privately, control downloads, hide ads and more …

OpenTalks.AI - Никита Андриянов, Анализ эффективности распознавания образов на нестандартных типах изображений на примере радиолокационных изображений местности и рентгеновских снимков багажа и ручной клади

99390fdb3165e59dcda59ab1b162fa1f?s=47 opentalks2
February 04, 2021

OpenTalks.AI - Никита Андриянов, Анализ эффективности распознавания образов на нестандартных типах изображений на примере радиолокационных изображений местности и рентгеновских снимков багажа и ручной клади

99390fdb3165e59dcda59ab1b162fa1f?s=128

opentalks2

February 04, 2021
Tweet

Transcript

  1. Nikita Andriyanov Analysis of the efficiency of pattern recognition on

    non- standard types of images: radar images and X-ray images of baggage and carry-on baggage
  2. Relevance Decoding of radar images Aviation Security 2

  3. 3 Data features -Registration in any weather conditions -Registration at

    any time of the day - X-ray scan Baggage on X-ray Image Earth’s Surface on Radar Image
  4. Radar Images Examples of radar images: (a) - plane, (b)

    - road, (c) - ship 4
  5. Original Dataset * Sandia National Laboratories Class Sample Type Train

    Validation Test Plane 9 2 3 Road 30 4 9 Ship 9 2 3 5
  6. Augmentation Examples of augmented radar images: (a) - plane, (b)

    - road, (c) - ship 6
  7. Augmented Sample Class Sample Type Train Validation Test Plane 189

    20 3 Road 210 28 9 Ship 189 20 3 7
  8. Neural Network Architecture 8

  9. Accuracy Layers, neurons Optimization, epochs Dataset Accuracy, % Train Validation

    Test 2x16+16 SGD, 30 Source 100 87,5 86,7 3x16+16 SGD, 30 Source 100 75 100 2x16+16 SGD, 30 Augmented 50 87,5 73,3 3x16+16 SGD, 30 Augmented 87,5 75 66,7 3x16+16 SGD, 80 Augmented 87,5 87,5 93,3 3x16+16 ADAM, 30 Source 87,5 87,5 93,3 3x16+16 ADAM, 30 Augmented 100 100 100 9
  10. Learning Learning characteristics: (a) - proportion of correct recognitions, (b)

    - loss function a b 10
  11. Results on Radar Images 11

  12. X-ray Images Examples of processed images of baggage and carry-on

    baggage 12
  13. 13 Neural Network Architecture

  14. 14 Original Dataset Sample Prohibited Permitted Train 895 1814 Validation

    62 107 Test 59 106 Total 1016 2027
  15. 15 Comparative Analysis Layers, neurons Dataset Epochs Optimization Accuracy CNN

    40-36-25 Grayscale 112x112x1 8 SGD 62.34% CNN 40-36-25 Color 112x112x3 8 SGD 66.48% CNN 64-48-32 Color 112x112x3 200 ADAM 69.32% CNN 40-36-25 Color 112x112x3 8 ADAM 69.62% CNN 64-48-32 Color 112x112x3 8 ADAM 75.32% CNN 64-48-32 Color 112x112x3 100 ADAM 81.75% CNN 64-48-32 Color 128x128x3 100 ADAM 85.71% CNN 128-128-32 Color 128x128x3 50 ADAM 90.47%
  16. 16 Learning

  17. 17 Results of processing

  18. 18 Best results

  19. 19 Accuracy dependencies Layers Accuracy, % 1 67.722 3 74.684

    5 77.848 7 82.278 Drop out Accuracy, % 0.1 79.114 0.2 77.215 0.3 62.025 0.4 33.544 0.5 32.911 Augmentation probability Accuracy, % 0.1 82.911 0.2 86.076 0.3 84.177 0.4 82.911 0.5 81.013
  20. 20 Conclusion 1) In this work, using convolutional neural networks,

    the problem of classifying radar images containing 3 types of objects was successfully solved: an airplane, a road, a ship. To improve the quality of the neural networks, a modification of the gradient descent algorithm and data augmentation were used. For most of the images, the network managed to achieve "confidence in its choice" over 90%. 2) Using the example of primitive modifications of the input data and the architecture of a neural network for X-ray scanning images of airport security services, a neural network was trained on a low-performance computer, which provided a result of more than 90% on a test sample. 3) The study of standard regularizations during training on the data of X-ray images showed general trends with the processing of optical images. Obviously, the accuracy in this case is directly proportional to the depth of the network. However, over time, a further increase in the number of layers will lead to a much larger increase in computational costs than a gain in recognition accuracy. Interestingly, resetting the weights in this problem led to a significant decrease in the recognition accuracy. At the same time, the maximum accuracy for augmentations was achieved at a probability of 0.2. At the same time, for the considered example, resetting the weights did not lead to an improvement in the quality of the network, and insignificant augmentation increased the accuracy by about 3.8%.
  21. 21 Acknowledgements The research was supported by the Grant of

    Russian Foundation for Basic Research and the Government of the Ulyanovsk Region, within the framework of Project No. 19-47-730011 Also special thanks to the organizers of the conference OpenTalks.AI-2021
  22. THANK YOU FOR YOUR ATTENTION! + 7 937 037 9690

    Nikita Andriyanov, PhD (candidate of technical sciences), Assistant Professor, Department of Data Analysis and Machine Learning, Financial University under the Government of the Russian Federation n.andriyanov nikita-and-nov@mail.ru