Color Story
This script detects the predominant colors in a given image. Since I intended it as an educational exercise in basic Computer Vision, I ignored existing OpenCV functions except for basic import and RGB <--> HSV data conversions. The script calculates average colors in HSV space (as opposed to RGB) in order to facilitate future color detection adjustments according to Hue groups.
Dependencies (virtual environment recommended):
conda install -c conda-forge opencv numpy matplotlib scipyTo use, run:
python color_detector_KMC.py <num_colors> <image_path>e.g.
python color_detector_KMC.py 5 images/flowers4.jpg[Input Image]
↓
┌─────────────────────────┐
│ Preprocessing │
│ ┌──────────────────┐ │
│ │ Downsample │ │
│ └────────┬─────────┘ │
│ ↓ │
│ ┌──────────────────┐ │
│ │ RGB → HSV │ │
│ │ Conversion │ │
│ └────────┬─────────┘ │
│ ↓ │
│ ┌──────────────────┐ │
│ │ Reshape to │ │
│ │ Pixel Array │ │
│ └────────┬─────────┘ │
└───────────┼─────────────┘
↓
┌─────────────────────────┐
│ K-Means Clustering │
│ ┌──────────────────┐ │
│ │ Initialize │ │
│ │ Centroids (FPS) │ │
│ └────────┬─────────┘ │
│ ↓ │
│ ┌──────────────────┐ │
│ │ Calculate │ │
│ │ Distances (HSV) │ │
│ └────────┬─────────┘ │
│ ↓ │
│ ┌──────────────────┐ │
│ │ Assign to │ │
│ │ Clusters │ │
│ └────────┬─────────┘ │
│ ↓ │
│ ┌──────────────────┐ │
│ │ Update Centroids │ │
│ │ (mean) │ │
│ └────────┬─────────┘ │
│ ↓ │
│ └──→ Iterate 10x │
└─────────────────────────┘
↓
[Sort by Hue]
↓
┌─────────────────────────┐
│ Generate Visualizations│
│ ├→ Color Swatches+Hex │
│ └→ Polar Plot (H vs S) │
└─────────────────────────┘
The optimization method is K-means clustering (KMC). I first selected an initiating set of pixel colors using farthest point sampling (FPS) in Hue space, selecting Hue values located at intervals of 180 / (# number of colors to detect). The initiating Saturation and Value are set to 122.5.
The program then compares each pixel to each initiating centroid, and calculates the distance betweeen them using the L2 norm in cylindrical coordinates, before assigning each pixel to the closest centroid. Keep in mind that for Hue, the values 0 and 179 are both red, which can potentially skew distance calculations in Hue space (e.g. 2 degrees is closer to 179 than to 6). To get around this, I implemented a wraparoud that takes the minimum of the difference in hues between the pixel and centroid.
IMPORTANT NOTE: OpenCV formats Hue values to range between 0 and 180. On the other hand, circular or cylindrical coordinate systems requre a range between 0 and 360 in radians. Therefore, Hue values are multiplied by 2 and converted to radians before the mean calculation, and then reconverted afterwards in keeping with OpenCV format.
To obtain the updated set of centroids, I calculated the HSV mean of each cluster. Since Hue space is circular, you need to use a circular mean, while Saturation and Values are averaged normally. You can read about the circular mean here: https://en.wikipedia.org/wiki/Circular_mean
The process repeats again for a hard-coded number of iterations (10 in this case, but you can change it by changing variable num_iter).
The program produces three figures as output:
The first shows the original image, alongside a resized version of the image to reduce the pixel load for calculation efficiency. The percentage of pixels used for the calculation is hardcoded to 1%, but can be changed in the script using variable 'perc'.
The second visualizes breakdown of these pixels in RGB and HSV space.
The third shows the distribution of common hues according to increasing Hue, along with the HEX codes of each color. Alongside this distribution we show these pixels plotted on a polar graph.
- Implement feature recognition to exclude photographed people from color stories (because that's just weird)
- Create a matrix plot of colors so that common colors of similar hues are grouped together