TensorFlow-Powered Vision For Pi-based robot

---

• TensorFlow Version: 1.0.1
• Written by Korean

Introduction

This is a Pi-based robot to implement visual recognition(Inception V3). The TensorFlow-Powered vision can recognize many objects such as people, car, bus, fruits, and so on.

• Hardware: Raspberry-Pi2, Sony PS3 Eye Camera

  (Available to use Logitech C270 USB camera with Raspberry Pi)

• Software: TensorFlow(v1.0.1), Jupyter-Notebook

• Schematics: Only USB Camera Part

My motivation

I was so curious about excellence of the image recognition with TensorFlow on Raspberry Pi. Also, the Jupyter notebook is very convenient to instantly code as a quick prototype. So, in terms of error rate of the image classification, Inception V3(3.46%) is more excellent than human(5.1%) whereas raspberry pi's processing speed is very slow compare to my laptop.

(Table: Jeff Dean's Keynote @Google Brain).

• Schematic diagram of Inception-v3

Requirements and Installation

• Install a webcam driver on your Rapsberry Pi.

sudo apt-get install fswebcam

• Test your webcam.

fswebcam test.jpg

• TensorFlow (V1.0.1): How To Install TensorFlow on Raspberry Pi

• Jupyter-Notebook: How To Install Jupyter-Notebook on Raspberry Pi

Quick Start

• You should install both TensorFlow(v1.0.1) and Jupyter notebook on your Raspberry Pi.

• First, clone the TensorFlow-Powered_Robot_Vision git repository here. This can be accomplished by:

cd /home/pi/Documents
git clone https://github.com/leehaesung/TensorFlow-Powered_Robot_Vision.git

next, cd into the newly created directory:

cd TensorFlow-Powered_Robot_Vision

Drive your jupyter notebook on your Raspberry Pi.

jupyter-notebook

The pre-trained data(inception_v3.ckpt) will automatically download when driving the Jupyter notebook. (Where: /pi/home/Documents/datasets/inception)

Source Codes

• TensorFlow-Powered_Robot_Vision.ipynb
• imagenet_class_names.txt
• inception_v3.ckpt (When implementing the jupyter notebook)

Results of Object Recognition

• Wow! The result is really awessome!!

How To Retrain

• This CodeLab(Google) is really useful in learning this stuff
• Train an Image Classifier with TensorFlow for Poets (YouTube)
• How to Retrain Inception's Final Layer for New Categories

References

• Rethinking the Inception Architecture for Computer Vision (Paper)
• Image Recognition::Tensorflow.org (Web)
• Hands-On Machine Learning with Scikit-Learn and TensorFlow (Book)
• Train your own image classifier with Inception in TensorFlow (Blog)
• "Large-Scale Deep Learning for Building Intelligent Computer Systems," a Keynote Presentation from Google
• How Convolutional Neural Networks work (YouTube)