Stephane Paquis
| IF No/Ng » 1 THEN SD/CBA class, ELSE PA/UTBA class
For each class, road surface texture is classified by analysing textural features extracted from residual distributions of
Y? .
multiresolution cooccurrence matrices. For each unknown texture, a parameters vector V € R^ is computed. Classifictigy
consists in comparing this features vector to a set of reference vectors. We compare Mahalanobis measures between
candidat and kernels simply.
5.2 Application
Our classification algorithm was applied to a set of 60 samples (15 per familly). These texture were digitized on tes
tracks of LCPC (Nantes) with a 768 x 572 spatial resolution, quantified on 256 grey levels. All manipulations were done
on dry pavement surfaces without distress features.
Each image was divided into 5 overlapping 512 x 512 subimages. Among them, the corresponding center one was used
as training set to estimate reference vectors nessecary to the second classification step. Remaining 4 subimages of each
texture, 240 subimages altogether, were taken as an unknown set to be classified.
The classification results are listed in table 5.2, which represents confusion matrix Con f(a;,œ;) where each entry gives
percentage of elements belonging to the class o; assigned to the class o ;.
CBA | SD | PA | UTBA
CBA 90 10 0 0
SD 0 97 3 0
PA 0 0 | 100 0
UTBA 0 0 0 100
Table 1: Classification results of 240 subimages from 4 x 15 road surface texture images.
Approach was also applied to a set of 200 new subimages from 4 x 10 images. We used the same training set as the one
defined previously. Classification results are shown in table 5.2.
CBA | SD | PA | UTBA
CBA 100 0 0 0
SD 0 88 | 12 0
PA 0 2 | 98 0
UTBA 0 0 0 100
Table 2: Classification results of 200 subimages from 4 x 10 new road surface texture images.
6 CONCLUSION
This paper has presented an image processing tool for achieving classification task of 4 categories of road surfaces textures.
Developped method is based on a non linear image transformation similar to a classical opening operator. This operation
allows to study image textures at several scales. Texture analysis step is based on information extracted from data of
a set of multiresolution cooccurrence matrix. The classification schem combines 2 complementary steps : a statistical
one, which values texture variation along the multi-scale decomposition, and a structural one, where a simplified texture
version is analysed geometrically.
First results of such a classification method are interesting and further work would consist in testing this approach ona
larger set of road surface texture images. Besides it woud be interesting to use this procedure to locate inhomogeneities
in texture in order to detect pavement surface defects.
REFERENCES
Chanda, B. and Majumder, D., 1988. A note of the use of the graylevel cooccurrence matrix in threshold value. Signal
Processing 15, pp. 149—167.
Chubb, C. and Yellot, J., 2000. Every discrete, finite image is uniquely determined by its dipole histogram. Vision
Research 40, pp. 485-492.
690 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
pe
Haddc
135.
Harali
Vision
Houze
de rec
Konik.
Traitei
Paquis
SPIE :
Rosen
Recog
Rouqu
champ
Rouqu
Une af
Segall.
8(2), p
Serra,
Toet, A
Treme:
segmei