See project on github!

Find-the-red-car

computer-vision-color-based-search

Special forces mobilize! Computer vision activate! Enhance!

In this exercise I use the python imaging library to locate red cars from aerial images of parking lots. To understand the surveillance state…you must become the surveillance state! Wait, that can’t be right…

Setup

Install python3 for your distribution.

$ sudo apt install python3

Install required python imaging library.

$ pip install pillow --upgrade

Usage

The script is run from command line in the form:

$ python main.py /path/to/img.jpg x_grid y_grid

For example, to process back-parking.jpg with a 5x5 grid run:

$ python main.py back-parking.jpg 5 5

Images of each tile are written to ./tiles . Each tile is a .jpg named in the pattern rank.coordinates.score.jpg. Where rank is the tile rank, from 0 = most-red and n = least-red. coordinates are coordinates of the tile within the original photo. score is a score indicating the amount of red present.

Explanation

0. The goal

The task given was to break the provided image into an nxn grid of tiles - then calculate the amount of the color red present in each tile.

The provided photo was of a parking lot with several red cars.

1. Creating tiles

The original image is broken into a grid of tiles. For example, 5x5 8x3 10x6 etc. Each tile has the same dimensions, determined by width x height of the original image.

Each tile is described by two coordinates…

a is the upper-left

b is the bottom-right

a  _______
  |       |
  |  tile |
  |       |
   ------- b

To fully describe a tile the following the coordinates of both a and b within the original image.

(a1, a2, b1, b2)

(0 ,0 ,100 ,100)

2. Scoring tiles

For each tile in the original image we need to calculate the amount of red that appears.

The “redness” of each pixel is calculated as:

redness = RED - GREEN - BLUE

For example, take this candy-apple red color (227,46,86)

redness = 227-46-86 = 95

The total score indicating the amount of red in the tile is the sum of the “redness” of each pixel. Tiles with much red will score higher, tiles without red will score lower.

The tiles are then ranked by their score. The 0th tile contains the MOST red and nth tile contains the LEAST red.

3. Analyzing results

Sifting through the ranked tiles - the program certainly seems to be identifying tiles by amount of red. However, to better communicate this the concept of focusing one last image is created.

Each tile is darkened/brightened by its rank. The tiles are then re-assembled into a final image showing visually how the program focuses on red areas of the image.

Original Image

Focused Image, coarse-grid

Focused Image, fine-grid

Another Image

Another Image, coarse-grid

Another Image, fine-grid