ashishd / vision

Vision

Vision

Computuer Vision and Image Processing with Python, OpenCV, and Keras.

---

🌁 1. Dehazing to Object Recognition 🚧

Human, and fire detection and recognition in indoor foggy environment.

i. Image Processing with OpenCV

A. Dehazing

a. Dark Channel Prior | GitHub | Paper

B. Object Recognition

a. Human detection with RPi | Github

b. HOG detectMultiScale | Blog (KR) | Blog (US)

ii. Image Processing with Deep Learning

📚 A. Dataset | Reference

a. RESIDE: V0 (REalistic Single Image DEhazing) | Homepage | Paper (IEEE) | Paper (arXiv)

b. I-HAZE | Homepage | Paper (arXiv)

c. D-HAZY | Homepage | Paper (IEEE)

d. O-HAZE | Paper (arXiv)

🏁 iii. Challenge

A. NTIRE2018 | Homepage

---

🎯 2. Vanishing Point to Control Posture ✔️

Using Image Processing with OpenCV

---

3. Depth to Obstacle Avoidance 🚧

i. Image Processing with Deep Learning

📚 A. Dataset

a. NYU Dataset V1 | Homepage | Paper

b. NYU Dataset V2 | Homepage | Paper

c. Sample Code for NYU

• How to read NYU mat file with python | Blog (KR)
  • Packages scikit-image matplotlib

> pip install scikit-image
> python -m pip install -U matplotlib

• Change Code for Windows or OS X | GitHub issue

import skimage.io as io

to

import matplotlib
matplotlib.use('TkAgg')
from skimage import io
io.use_plugin('matplotlib')

c. RGBD Dataset | Homepage

B. Model

a. FCRN | Paper (arXiv) | GitHub

---

4. Optical Character Recognition to Path Planning 🚧

i. Tesseract | GitHub | Demo

ii. PyTesseract | GitHub | Install with pip

> pip install pytesseract

A. Install Windows Version | GitHub | Download

B. Tesseract training | GitHub

C. Variable-size Graph Specification Language (VGSL) | GitHub

D. StreetView Tensorflow Recurrent End-to-End Transcription (STREET) | GitHub

E. Korean OCR | Blog (KR)

• a. Remove spaces | Blog (KR)
• b. Once you change the route, you need to turn off and restart Pycharm.
• c. Remove special characters | Blog (KR)
• d. Tesseract Optimal conditions | Blog (KR)
• e. Color Reversal | Blog (KR)

image = cv2.bitwise_not(input_image)

• f. Resize | Blog (KR)

image = cv2.resize(input_image, dsize=(0, 0), fx=0.3, fy=0.7, interpolation=cv2.INTER_LINEAR)

🏁 iii. Challenge

A. Robust Reading Competition ICDAR | Homepage

Overview - ICDAR 2019 Robust Reading Challenge on Multi-lingual scene text detection and recognition | Homepage

iv. Korean STR

Korean Scene Text Recognition by Character-level

📚 A. Dataset

a. Dataset Version 1 (.png) | MNIST.zip | PHD08.zip

Dataset	Language	# of fonts	# of characters	total # of images
PHD08	Korean	9	2,350	5,139,450
EMNIST (ByClass)	Number & English	Hand Writing	62	814,255

b. Dataset Version 2 (.png) | 7z

Language	# of fonts	# of characters	total # of images
Korean	70	11,172	782,040
Number	70	10	700
English	70	52 (upper 26, lower 26)	3,640

c. Dataset Version 3 (.png) | 7z | zip

Language	# of fonts	# of characters	total # of images
Korean	69	972	58,320
Number	69	10	690
English	69	52 (upper 26, lower 26)	3,588

d. Dataset Version 4 (.png) | 7z | zip | pkl

Language	# of fonts	# of characters	total # of images
Korean	69	972	116,640
Number	69	10	1,240
English	69	26	3,588

• Korean and numbers use two font sizes.
• Incorporate uppercase letters into lowercase letters.
• Font sizes of Korean and number are 44, and 54, font size of English is 44.
• Image size is 64.
• Hangul characters refer to this and calculated the frequency.

---

Install

i. Install OpenCV on Computer

ii. Install OpenCV on Raspberry Pi

iii. Install TensorFlow CPU on Computer

iv. Install TensorFlow GPU on Computer

v. Install & Uninstall TensorFlow on Raspberry Pi

vi. Install Keras & Keras Reinforcement Learning

vii. Install Pytorch

Viii. Install Theano

iX. Virtual Environment with Conda

X. Setup Ubuntu

---

• Image Processing vs. Computer Vision Reference
• Dehazing - Image Processing/Object Recognition - Computer Vision
• Vanishing Point - Computer Vision
• Depth Prediction - Computer Vision
• Opitcal Character Recognition - Computer Vision