TV Recognizer

Brazil Tv Station Logo Recognizer

This project was originally created to be part of Beep App project. The work uses Flutter framework to process a TensorFlow Lite model (FlatBuffer format) containing a Trained Convolutional Neural Network (CNN). The model processes a live video from Smartphone Camera (Using Camera plugin 0.5.3) on FLUTTER in order to recognize Tv Station Logos.

My ARTICLE on Medium about this logo detection: MEDIUM ARTICLE LINK

Goal:

This basically recognize 6 main Brazil Tv Stations Logos (SBT, Globo, Record, RedeTV, Band, TvBrasil) on Android Smartphone and PC Webcam.

Data:

• TFLITE file (TensorFlow Lite) on /assets folder, was constructed using:

  • Darknet: Open source neural network framework (C and CUDA), for GPU Computation LINK
  • YOLO V2 Tiny: Real-time object detection system, V2 Tiny version due the compatibility with TensorFlow Lite.
  • Fastai: vision library to manipulate, augment and also check the results of training.
  • Darkflow: To convert to Tflite PB (Protobuf) containing the model' weights and the graph definition.

• The Notebook Logo_Data_Augmentation was created on Google Colaboratory and show the whole code and explain the Data Augmentation Process.

• The Flutter Folders (TvLogo_Detect) consists of the Distance detection (for entire Television), and Close detection (for just logo on entire phone screen).

• Pre-made APKs to put directly into smartphone.

Setting Up:

First you will need to install Flutter for your platform: https://flutter.dev/docs/get-started/install

Then run flutter doctor and be sure that no issues are reported for the sections relevant to your platform.

In order to run this ANDROID APP on a smartphone just do flutter run in the TVLogo Far or TVLogo Close directory.

Install from pre-made APK:

You can install the apk directly on Smartphone:

• APK for Far distance: LINK
• APK for Close distance: LINK

PREVIEW:

Detection on Webcam:

Detection on Smartphone:

Detecion on Entire Screen:

Copyright © 2019 leoitcode