Color segmentation of still images


These segmentation techniques were elaborated, to solve multimedia issues in the large context of indexing, but this time for still images, and in the framework of several contracts, with the following partners:

Center of Mathematical Morphology (Ecole des Mines de Paris).
CCETT (France TÚlÚcom).

Back to top


The proposed segmentation presents a specific feature: one searches only to extract large homogeneous regions in color and texture, as in a first indexing process small details are not absolutely necessary to extract.

Besides this typical characteristic, our segmentation offers to additional new features:

the enhanced HSV transform,
the unusual succession of pre-treatments.
The proposed segmentation, and particularly the HSV transform step, finds its applications in the indexing of images, but also in problems of object classification, people detection and recognition, etc.
Back to top

Enhanced HSV transform

Back to top

The aim of this transformation, especialy developped for our segmentation, is to find a pixel classification allowing to separate the pixels with a significant color from the pixels for which their color is not obvious and rather similar to a grey level. Once such a classification is achieved, one proceeds to an enhancement of the "real" colors and of the grey levels.

The chosen color space is the HSV space: we split this space into two distinct regions.
The limit between the two classes is automatically determined, but figure 1 gives several classifications according to different chosen thresholds.
The automatic splitting realized by the HSV transform corresponds to the second image of the second line (Fig.1).

Fig. HSV space Fig. HSV splitting into two regions.
Back to top of HSV transform

Fig.1 Examples of the HSV tranforms for increasing threshold values. The original image is on the top-right.
Back to top of HSV transform

By its action of enhancement of the difference between pixels with a color and pixels with a grey level, this transform constitutes an interesting pre-treatment for numerous applications. We illustrate its abilities in an example of people detection (Fig.2). In this example, the HSV transform is simply followed by a thresholding of the hue values in the image.

Fig.2 Application of face detection. We present successively the original image, the HSV transform,
with automatic pixel classification, and the result of the hue threshold.
Back to top of HSV tranform

Back to top

Main steps of the color segmentation

Back to top

More than in each individual step, the original aspect of the segmentation consists in their succession:

Original image
Enhanced HSV transform
Color reduction
Vectorial median filtering
Color gradient
Hierarchical watershed

We will not come back to the HSV transform. The next step of color reduction has been chosen in order to reduce drastically the amount of information, as only 16 colors are kept in the end. In spite of the loss of information, particularly for the small details, this does not affect the large regions of the image which we want to extract.

Fig.3 Reduction to 16 colors. The original image is on the left and the result is presented on the right.

In order to avoid the appearing of false colors, due to a filtering of each color component separately, a vectorial median filter is then applied. The chosen color gradient is a mixing of both Lab and HSV gradients. Finally, we apply the classical algorithm of hierarchical segmentation by watershed. After one or two steps of segmentation the result is conform to the desired aim: large homogeneous regions only have been kept.
Fig.4 Vectorial median filter. Fig.5 Color gradient
(mixing of Lab and HSV gradients).
Back to the segmentation steps

Fig.6 Steps of the hierarchical segmentation. We present successively the original image, the original watershed
and the two first steps of the hierarchical segmentation.

Back to top

All the original images in this page come from TV newscasts of the french TV channels TF1, FR3, A2 or M6 and are therefore copyrighted by these channels. All other images or photos are copyrighted by the CMM. These documents are protected by the law on the author rights, each non-authorized copy or use is therefore strickly forbidden.
Last update : 01 - 06 - 99