peroon / HelloProjectRecognizer

Idol Recognizer (using Deep Learning) and Viewer

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Hello! Project Recognizer

Goals

  • Recognize idol in video using Deep Learning
  • YouTube viewer using the result information

Viewer Demo

Official Page

Hello! Project

Groups

How to prepare training images

Get images

  • Google Image Search
  • Bing Image Search
  • Youtube & Face Detection

Crop faces

  • refer to face.py
  • face detection using dlib
  • detected faces are placed in the candidates directory

Sort by hand

  • candidates are sorted by hand and placed in recognized idol directory

Eliminate by hand

  • Remove almost same image and blurred image
  • This prevents the number of training data from increasing unnecessarily
  • TODO: Efficiency can be improved if there is a tool to sort by image similarity

Classifier

  • Keras Resnet
  • MXNet Resnext

What can this framework do?

  • Idol face recognition
    • Training
    • Prediction
  • Visualization
    • using t-SNE
    • face similarity between idols

Accuracy

  • 0.627 on validation set (3000 images)
    • 0.9 is desirable and I think it is possible.

Environment Settings

How to collect training images

Google Search API

How to add a video to list

  • Download mp4 by running code/collect/youtube_downloader.py
  • Run code/video_analyzer.py after assign youtube id
  • Run docs\view\list\list_data_maker.py

TODO

  • Since it is meaningless to learn almost the same face image, we want to eliminate overlapped images by sorting by image similarity
  • Make READMEs ogranized
  • How to manage the change of group of idol or resignation
    • Withdrawal while maintaining a good relationship
      • Visible?
    • Withdrawal due to causing problems
      • Invisible?

About

Idol Recognizer (using Deep Learning) and Viewer


Languages

Language:Python 95.7%Language:Jupyter Notebook 4.3%