A method of compressing pictures---machine learning's K-means

Background Description:

RGB encoding: For an image that directly uses 24bit for each pixel, each pixel uses an 8-bit unsigned integer (0-255) to denote red or green or blue.

Compression Purpose:

The 128x128 size picture is represented by the original 24bit-compressed into->16bit to represent the image of each pixel.

Compression method:

For each pixel, use K-means to select 16bits to represent the original 24bits. Of course, the original 24bits is expressed by calculating the cluster of 16bits size per pixel space.

Implementation steps:

1. Read the original 128x128-sized picture into a 3-D matrix A. For example, a (50, 33, 3) indicates that the original image is RowNum = Columnnum = the b dimension in RGB is the corresponding

Color value (0-255). So we get an X = (M X 3) matrix of Piexl colors (where m = 128 * 128 = 16384).

2. Run the K-means algorithm, take k= 16, so that each pixel cluster to a cluster K (the choice about the initial K-point of K-means is randomly selected in X as the centroid of K points). Use K, the location of the centroid point, to represent the original pixel value. Here k=16, we can use four bits to represent the possibility of 16 classifications. Then you need 4bits to record the compressed classification of the current PIEXL. (Of course, there will be a table K (16 colors). Why not 4bits to map, 2 binary conversion is too cumbersome! The direct mapping here), RGB (24bits)).

Compression effect:

Pre-Compression size: 128*128*24 = 393216 bits;

Compressed size: 128*128*4 + 16*24 = 65920 bits;

Compression factor is close to 6!

The code is in Gitlab machine learning ex7.

Core ideas:

K-means image compression, compression is not a color. The original 24-bit color, now also used 24-bit color. But the coding changed, turning the original 24bits color into a 16bits color. And this process is the use of K-means clustering. Then the distinction between colors becomes less obvious, resulting in reduced resolution.

Has fun, good luck!

And this are my picture Kebe compressed:

A method of compressing pictures---machine learning's K-means