Before giving details of the formal method we used to evaluate our proposal, we would like to emphasize a feature that we believe is an important contribution of our proposal: the use of a combination of colour and texture properties. To illustrate this, Figure 3 shows a simple experiment which consists of the segmentation of a mosaic composed by four regions, each of which has a common property, colour or texture, with their adjacent regions. Specifically, each region has the same colour as its horizontal neighbouring region, while the vertical neighbouring region has the same texture. As is stated in the first two examples of Figure 3, the method allows the colour texture properties to be modelled and the four regions to be correctly segmented. On the other hand, we included in Figure 3 a third experiment which consists on the segmentation by using only colour information. Third image shows the smoothed version of the first mosaic image, and clearly illustrates that the original image contains only two colours. Therefore, the segmentation of the image using colour information, although different techniques can be used (considering a Gaussian distribution on the smoothed image, modelling using a mixture of Gaussians on the original image, or other techniques such as the kernel density estimation), will only allow us to identify two colour regions and it is not possible to distinguish regions with the same colour but different texture. As is stated, in some cases colour alone does not provide enough information to perform colour texture analysis, and in order to correctly describe the colour texture of a region, we need to consider not just the colour of pixels, but the relationships between them.
|
The described segmentation method can be performed over any set of
textural features. The result of comparing the relative merits of
the different types of features have been nonconclusive and an
appropriated set of features has not emerged in all cases. For the
experimental trials shown in this article we used the
co-occurrence matrices proposed by Haralick et
al. [24]. Two of the most typical features, contrast
and homogeneity, are computed for distance and for
,
,
and 135
orientations to constitute a 8-dimensional feature vector.
Moreover, the (
) colour space has been chosen to
model the colour.
In order to evaluate the proposed colour texture segmentation
technique, we created mosaic images by assembling
subimages of size
of textures from the VisTex
natural scene collection by MIT (
http://www-white.media.mit.
edu/vismod/imagery/VisionTexture/vistex.html), which we have
called from to
. Furthermore, we added
mosaics
,
and
, provided by Dubuisson-Jolly and Gupta which were
used to evaluate their proposal on colour texture segmentation
described in [8]. A subset of colour texture mosaic
images with obtained segmentation results is shown in
Figure 4.
|
The evaluation of image segmentation is performed by comparing each result with its ground truth and recording the error. Specifically, we use both region-based and boundary-based performance evaluation schemes [25] to measure the quality of a segmentation. Region-based scheme evaluates the segmentation by measuring the percentage of not-correctly segmented pixels considering the segmentation as a multi-class classification problem. Meanwhile, boundary-based scheme evaluates the quality of the extracted region boundaries by measuring the distance from ground truth to the estimated boundary.
Images were processed by our segmentation algorithm using various
set of parameter values for the weight of colour (parameter
) and texture information, as well as the relative
relevance of region (parameter
) and boundary information
in the segmentation process, and best results have been obtained
with
and
. Note that a predominant role
is given to colour and region information. Table 1
shows the quantitative evaluation of results obtained using this
parameters setting over the set of mosaic images. Summarising, a
mean error of
has been obtained in the region-based
evaluation for the whole set of test images. While the mean error
at the boundary has been of
pixels. Furthermore, our
proposal obtained errors of
,
and
in
the segmentation of
,
and
, respectively (see
segmentation results of these mosaic images in second row of
Figure 4), which can be compared to the segmentation
results shown in the work Dubuisson-Jolly and
Gupta [8]. Their proposal is a supervised
segmentation algorithm based on the fusion of colour and texture
segmentations obtained independently. Both segmentations are fused
based on the confidence of the classifier in reaching a particular
decision. In other words, the final classification of each pixel
is based on the decision (from colour or texture) which has
obtained a higher confidence. Our results have to be considered as
very positive since they significantly improve colour texture
segmentation results presented in [8].
Region-based | Boundary-based | |
(% error) | (pixels distance) | |
M1 | 2.207 | 0.352 |
M2 | 0.280 | 0.145 |
M3 | 0.731 | 0.237 |
M4 | 2.375 | 0.588 |
M5 | 1.663 | 0.786 |
M6 | 2.352 | 0.341 |
M7 | 1.451 | 0.596 |
M8 | 6.344 | 1.774 |
M9 | 3.609 | 3.430 |
M10 | 0.095 | 0.028 |
M11 | 3.550 | 0.962 |
M12 | 1.955 | 0.852 |
Mean | 2.218 | 0.841 |
Std | 1.711 | 0.940 |
Furthermore, the performance of our proposal for colour texture segmentation has been finally tested over a set of real images. Natural scenes predominate among these images, since nature is the most complex and rich source of colour and textures. Some colour texture segmentation results are shown in Figure 5. Meaningful regions in images are successfully detected and the usefulness of our proposal for colour texture segmentation is demonstrated. Furthermore, we want to emphasize some aspects related to the obtained results. See the last example of Figure 5 which shows the segmentation of a monkey among some leaves. The monkey is correctly segmented and, moreover, although the animal is absolutely black several parts of its skin are identified due to their different textural properties. Similar situations occur with other images in which animals are present. In the image with a leopard, region at neck which is not composed by typical spots of the animal, is detected and the same occurs with the lizard image in which the body of the animal, neck and belly are segmented as different regions. It is true that in these cases many of human would group all these regions to compose a single region related to the whole animal body. Nevertheless, this process of assembling is more related to the knowledge that we have about animals that to the basic process of segmentation. Hence, we believe that the segmentation performed by our proposal is correct as it distinguishes regions with different colour texture. The task of region grouping, if necessary, should be carried out by a posterior process which uses higher-level knowledge.
The correctness of boundaries obtained in these segmentations is also shown by the sketch of detected borders over original images. As has been pointed out, texture segmentation is specially difficult at boundaries and great errors are often produced at them. Hence, we want to note the accuracy of segmentations considering not only the correct detection of regions, but also the precise localisation of boundaries between adjacent textures.
|