1 of 26

Birds Species Classification �with Deep Learning

Daria Grigoreva

Viktor Krasnov

Ivan Kovalev

Mikhail Detiuchenko

Tatyana Ulybysheva

​

Prepared by:

2 of 26

Contents

  • Problem
  • High-level design:
    • Use Case
    • Swim lane diagram
    • IDEF0
    • Digital twin
    • Knowledge base
  • CNN
  • Application

3 of 26

Problem

Global warming has affected local ecosystems

Birds that usually do not migrate now are found

beyond their natural habitat areas�

Detecting migrating tropical birds

with the help of cameras

would help to find changes in the

ecosystems and take measures

​

For that, we need a application

to quickly determine

the species of the found birds

​

4 of 26

Solution: application development

5 of 26

Solution: application development

6 of 26

Project’s Swim-lane diagram

7 of 26

Application Aims

  • Scientific research: tracking birds’ habitat
  • Ecological intervention: finding ecosystems’

disbalance and protecting species

  • Education: training of students
  • Open data (for all interested users)

8 of 26

The key element of the application is the Image Recognition Model

9 of 26

Image Recognition Model: IDEF0

10 of 26

Similar projects

  • ProtoPNet of Duke University
  • BirdNET - analysis of audio recordings in a mobile app

11 of 26

Digital twin

Creating a digital twin will allow to analyse:

  • possible users of the application
  • application load depending on the season and region of use
  • the volume and cost of the hardware infrastructure
  • optimal steps for software development
  • prospects for expanding the functionality and scope of the application

12 of 26

Knowledge base

Knowledge base provides the structure of the model

It is connected to the model through the variable — number of epochs

13 of 26

Tools and resources

  • Kaggle dataset
  • Knowledge base
  • GPU and Google Collab
  • Python libraries
  • PyQT
  • Qt Creator
  • GNU C++ Compiler
  • Jupyter Notebook
  • Sublime Text

14 of 26

Dataset

  • 450 bird species
  • 70 626 training images (5 images per species)

15 of 26

CNN

The best model for image classification is Convolutional Neural Network

​

16 of 26

17 of 26

18 of 26

19 of 26

20 of 26

21 of 26

22 of 26

23 of 26

24 of 26

25 of 26

Multi-agent system: client-server architecture

26 of 26

Other ways to apply the project

  • Training the model for different animals
  • Using other data from the cameras
  • Commercial use of the application (gaming)
  • Social media