ColorVision

1
ColorVision

• Koushik Krishnan
• Ajit Karthik Mylavarapu
• Ramanan Raghuraman
• Karthik Vijayraghavan

2
Outline

• Algorithm overview
• Object identification and color correction
• Color calibration and classification
• Implementation
• Limitations
• Demo

3
Algorithm Overview
4
Algorithm Overview
Calculate average RGB value for the white index
card
Calculate ScaleFactor for the lighting condition
Capture image in the cellphone
Scale the non-white pixels in the image
Bin the pixels and select bin with maximum value
5
Reference Card and Object Identification

• Capture image using the phone
• Bin all pixels based on average intensity
• Highest intensity bin used as approximation for
  reference card intensity
• Traverse image and separate reference card from
  object
• Find average R, G and B values of reference card
  pixels, and use it to find Scale Factors
• Adopted from Gray World Algorithm

6
Color Correction

• Multiply each pixel in the array of object pixels
  with ScaleFactor

indoors on a bright day with blinds opened
outdoors on a bright day in the shade
7
Algorithm Overview
Calculate average RGB value for the white index
card
Calculate ScaleFactor for the lighting condition
Capture image in the cellphone
Scale the non-white pixels in the image
Bin the pixels and select bin with maximum value
8
Need for Color Calibration
R G B
R G B
R G B
Color Correction
Color 2
Color 1

• Camera white balance and auto exposure are
  unknowns of the system.
• The color after the color correction step may not
  look like how the color appears in the real world

9
Calibration Examples
Color in Real world
Color in Phone
Color after scaling
10
Binning Classification
Input Color
Bins
Colors

• Divide entire color space in closely spaced bins
  (244)
• Club multiple bins into single colors (41)
• Perceptually two colors may be similar but may be
  far apart in RGB color space

11
Color Classification
12
Implementation

• Algorithm developed in MATLAB
• But, K.I.S.S
• Symbian C vs Java
• Java is more portable
• Faster development with Java
• Java is less powerful
• Implementation Considerations
• Use downsampled image (80 x 60)
• Simple Math
• Pre-recorded sounds

13
Limitations

• Camera
• Auto white balancing
• Gamma correction
• Poor quality in low-light conditions
• Java ME API does not allow control over the
  flash/exposure/white balance settings
• Algorithm
• Shades of gray and very light colors
• Dark colors close to black

14
Dropped Ideas

• Using CIELab Color Space
• Using the phones internal white-balancing
• Taking 2 separate pictures of reference card and
  object
• Using a Black-edged reference card to make
  object-separation easier

15
Conclusions

• ColorVision detects up to 41 unique colors and
  announces it to the user
• Cell phone camera does automatic color balancing
  and exposure control
• ColorVision counters the cell phone color
  balancing by use of a reference white card
• Fast response from the phone requires the use of
  simple mathematical operations
• ColorVision does not use complex mathematical
  operations

16
Thank You for listening!