We have previously studied the approaches which permit to associate a texture to an image, and some properties of such transformations. To do so, we use image transformations which act on the pixels of the original image. Some of these transformations have the following property: they statistically mix in a very homogeneous way the pixels of the image. We say that such a transformation is a Quasi-Mixing Transformation (or QM-transformation for short).For an image, with the use of a QM-transformation we obtain a locally very homogeneous spatial representation of this image (which was previously not homogeneous). This representation appears to be a texture. It is possible to define criteria which permit to quantify the local homogeneity degree of this texture in order, for example, to extract some statistical attributes from a reduced part of the image.We show that the statistical model of the texture that we obtain allows to define a representation space in order to discriminate the texture classes and thus the colour images.