Experiment results

We have conducted retrieval tests both on texture grey value images and real world colour images. The texture database used in the experiments consists of a set of 116 Brodatz different texture classes. The set includes different categories of textures in the Brodatz album such as regular textures, oriented textures, natural textures and bidirectional textures. Each of the

512 × 512

images is divided into 16

128

128 ×

overlapping sub-images. The real world images consist of a set of nearly 2500 photographic images.

We used the monomial kernel function,

^{M X} ( ) ^{= X} ( ) ( ) ^{1, 0 ⋅ X 0, 3}

^to

extract the colour features from the three levels of RGB colour space, and the 3D

8 8

8 × ×

colour histogram was applied to the intermediate image.

We used Gabor filters as a group of wavelets on the image with a different ori-entation (K=6) at different scales (S=5) and

U

_l

= 0 . 05

,

U

_h

= 0 . 4

, s and t range from 0 to 128, i.e., filter mask size is

128× 128

, to obtain the texture features. The texture feature components are the first and second moments of energy that are ext-racted from each channel, as in (14).

A query image is any one of the images in the database. This query image is then processed to compute the texture feature vector and the colour

8 × 8 × 8

histogram and is compared with the respective features of all the images in the database.

To compare the texture feature vectors of query image Q and a target image T in the database, we used

( ) ⁼ ∑∑ ( )

( ) (

² mn^T

)

²

and to compare the colour histogram we used the well known statistical method

χ

²

to determine whether two distributions differ:

( ) ⁼ ∑ ( ⁻ ₊ )

Let Dt be the result of the difference between the query image and a database based on texture feature vectors, and Dc is the result of the difference between the query image and a database based on a colour histogram. The total difference thus takes the form:

Dc Dt

D

_tc

= α ⋅ + β ⋅

, Where

α + β = 1

Figure 1 gives a retrieval result from the texture image database. All the 20 simi-lar textures in the database are retrieved in the first 20 images, and the remaining ones are also relevant. Figure 2 shows a retrieval result from a real world image database. As can be seen, most of the retrieved images are images with similar textu-res and colours to that in the query image. It shows us that the proposed technique works well for retrieval of images with homogeneous colour and texture overall.

Figures 3 and 4 displays the comparison of the average recalls of the query images taken from texture image database and real world image database respecti-vely, the recall for which is defined as:

images

In this paper we have used the nonlinear monomial kernel function to capture the local structures of image content to find out the colour features. Such a technique does not require any segmentation and therefore works fully automatically [9, 12].

We have also used the Gabor function by applying some of appropriate dilations and rotations to base the texture features of the image on.

The extracted texture and colour features are then used to measure the similariti-es between the query image and the image database by weighting thsimilariti-ese feature typsimilariti-es according to needs or according to image type. The retrieval results have been shown and examined.

Figure 1: Image retrieval results from a texture image database

Figure 2: Image retrieval results from a real world image database

Figure 3: Comparison of the average recalls by using different features: a solid curve for the combination of colour and texture features, a dashed curve for texture

feature and a dotted curve for the colour feature

Figure 4: Comparison of the average recalls by using different features: a solid curve for the combination of colour and texture features, a dashed curve for texture

feature and a dotted curve for the colour feature

References

[1] H. Burkhardt and S. Siggelkow, “Invariant features in pattern recognition fun-damentals and applications. In I.Pitas and C. Kotropoulos, editors, Non-linear Model Based Image/Video Processing and Analysis, pages 269–

307. John Wiley & Sons, 2001.

[2] George M. Haley and B. S. Manjunath, “Rotation invariant texture classification using a complete space-frequency model,” IEEE transactions on Image Processing, vol. 8, no. 2, pp. 255–269, February 1999.

[3] B. S. Manjunath and W. Y. Ma. “Texture features for browsing and retrieval for large image data” IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 18, NO. 8, August 1996, pp. 837–842.

[4] B. S. Manjunath, P. Wu, S. Newsam, H. D. Shin, “A texture descriptor for browsing and similarity retrieval” Signal Processing: Image Commu-nication, 2000.

[5] R. Manthalkar, P. K. Biswas, “Rotation invariant texture classification using Gabor wavelets” Asian Conference on Computer Vision, pp. 493–498, Jan. 23–25, 2002, Melbourne Australia.

[6] R. Manthalkar, P. K. Biswas, “Color Texture Segmentation using multichannel filtering” Conference on Digital Image Computing Techniques and App-lication, pp. 346–351, Jan 21–22, Melbourne, Australia, 2002.

[7] R. Manthalkar, P. K. Biswas, B. N. Chatterji, ”Rotation and scale invariant textu-re classification using Gabor wavelets“ second international workshop on texture analysis and synthesis, 2002 Copenhagen.

[8] H. Schulz-Mirbach, “Invariant features for gray scale images” In G. Sagerer, S.

Posch, and F. Kummert, editors, 17. DAGM-Symposium, pages 1–14, Bielefeld, Germany, September 1995.

[9] S. Siggelkow and H. Burkhardt, “Image retrieval based on colour and nonlinear texture invariants” In S. Marshall, N. Harvey, and D. Shah, editors, Pro-cessings of the Noblesse Workshop on Non-Linear Model Based Image Analysis, pages 217–224, Glasgow, United Kingdom, July 1998.

[10] S. Siggelkow and H. Burkhardt, “Image retrieval based on local invariant featu-res” In IASTED International Conference on Signal and Image Proces-sing (SIP), pages 369–373, Las Vegas, NV, October 1998.

[11] S. Siggelkow, M. Schael, and H. Burkhardt “SIMBA- Search IMages By Appe-rance” In the B. Radig and S. Florczyk, editors, Pattern Recognition, DAGM, LNCS 2191, pages 9–16, München, Germany, September 2001.

[12] Sven Siggelkow. “Feature Histograms for Content-based Image Retrieval”

Ph.D thesis, Albert-Ludwigs-Universität Freiburg, Germany, 2002.

In document AGRIA MEDIA 2004 (Pldal 171-176)