Dept. of Information Systems
Hansung University Seoul, Korea, 136-792
So Hee Jeon
Dept. of Earth Science Education
Seoul National University
Seoul, Korea, 151-742
Dept. of Earth Science Education
Seoul National University
Seoul, Korea, 151-742
In general, texture analysis approaches are used for recognition and distinction of different spatial characteristics of spatial arrangement and frequency of tonal variation related to patterns or phenomena contained in the digital image or the sensor image. Previous works related to texture image have been carried out into the three categories: development and improvement of texture extraction algorithms, comparison between texture extraction schemes, and domain application of extracted texture images. These types of researches are similar to other cases in digital image processing, such as image classification. Some previous researches compared texture analysis methods; Dulyakarn et al. (2000) compared each texture image from GLCM and Fourier spectra, in the classification. Maillard (2003) performed comparison works bewteen GLCM, semi-variogram, and Fourier spectra at the same purpose. Bharati et al. (2004) studied comparison work of GLCM, wavelet texture analysis, and multivariate statistical analysis based on PCA (Principle Component Analysis). In those works, GLCM is suggested as the effective texture analysis schemes.
Considered application of texture images, these secondary images can be utilized to classification with multi-spectral data as an additional layer or layers. Zhang (1999) combined multi-spectral classification and texture filtering for building detection in the urban area, and suggested that this approach increases classification accuracy. On the other hand, Smith et al. (2002) said that texture image is not always good to accuracy of quality in the multi-spectral classification. In the urban remote sensing, texture image analysis is one of useful approaches for urban class extraction and separation in Wang and Hanson (2001) Herold et al. (2003). As for useful types of texture image by GLCM, Franklin et al. (2001) and Kiema (2002) proposed that homogeneity is the most useful one among several types of texture measures. In the case of GLCM algorithms, new algorithms related to performance improvement have been proposed by Al-Janobi (2001) and Clausi and Zhao (2003), in the fast computation aspect.
This study and implementation concerned is based on the original concept of GLCM (Grey Level Co-occurrence Matrix) and GLDV (Grey Level Difference Vector), which are the most popular texture image generation and analysis scheme, summarized by Haralick et al. (1973), Parker (1997) and Hall-Beyer (2004). Various types of extracted texture image were investigated and then HIS data fusion with those also applied to 1 M Pan-sharpened IKONOS image at the urban area composed of complex features.
2. Brief Overview of Texture Analysis: GLCM and GLDV
Basic of GLCM Texture considers the relation between two neighboring pixels in one offset, as the second order texture. The grey value relationships in a target are transformed into the co-occurrence matrix space by a given kernel mask such as 3*3, 5*5, 7*7 and so forth.
In the transformation from the image space into the co-occurrence matrix space, the neighboring pixels in one or some of the eight defined directions can be used; normally, four direction such as 0°, 45°, 90°, and 135° is initially regarded, and its reverse direction (negative direction) can be also counted into account.
Therefore, general GLCM texture measure is dependent upon kernel size and directionality, and known measures such as contrast, entropy, energy, dissimilarity, angular second moment (ASM) and homogeneity are expressed as follows:
where i and j are coordinates of the co-occurrence matrix space, g(i,j) is element in the co-occurrence matrix at the coordinates i and j, Ng is dimension of the co-occurrence matrix, as grey value range of the input image. While, in GLCM texture measure, normalization of GLCM matrix, by each value dividing by the sum of element values, is applied, and then g(i,j) is replaced to the probability value. Furthermore, measures related to each texture variables also can use weights related to the distance from the GLCM diagonal.
As for GLDV texture measures, it is the sum of the diagonals of the GLCM mentioned above. In GLDV, these types of texture measures are also possible by Hall-Beyer (2004).
Texture measure in GLCM and GLDV needs to interpretation reference. Each measure contains different meaning for this. Homogeneity is measure for uniformity of co-occurrence matrix, and if most elements lie on the main diagonal, its value will be large, compared to other case. Dissimilarity measures how different elements of the co-occurrence matrix are from each other. Contrast measures how most elements do not lie on the main diagonal. Entrophy is to measure randomness, and it will be the maximum when the all elements of the co-occurrence matrix are same. In case of Energy and ASM, they measure extent of pixel pair repetitions and pixel orderliness, respectively.
3. Implementation: GLCM and GLDV
Stand-alone application program for GLCM and GLDV texture measure and texture image creation is implemented in this study (Fig. 1). In this program, general graphic image formatted as jpp, tiff, bmp can be used as input data. The menu functions for user selection are as follows:
Set_Depth_Level: Grey level quantization into 2 (Binary), 8, and 16
GLCM_Texture GLDV_Texture: Computation and generation of six types of texture measures of Homogeneity, Dissimilarity, Contrast, Entrophy, Energy, ASM(Angular Second Moment)
While, a user determines two texture parameters such as window kernel size and direction in the main frame. After the grey value relationships in a target are transformed into the co-occurrence matrix space by a given kernel mask such as 3*3, 5*5, 7*7 and 11*11, the neighboring pixels as one of the four directions as East-West of 0°, North-East of 45°, North-South of 90°, North-West of 135°, and omni-direction will be computed in the co-occurrence matrix space. Among them, texture image of omni-direction is obtained as the average value with those of four directions.
Fig. 1. GLCM and GLDV texture image generation application program.
4. Texture Analysis and HIS Fusion of Texture Images
Several texture images by GLCM and GLDV are presented and investigated, and sample image in Fig. 2 is from Demin (2002).
Fig. 2. Test image (excerpted from Demin, 2002).
Fig. 3 demonstrates direction dependency of texture measures. In Fig. 3, the notation of (8,GLCM, 5*5, E-W, ENT) means quantization level (2,8, or 16), application scheme (GLCM or GLDV), rectangle kernel size (3, 5, 7, or 11), direction (EW, NE, NW, NS, or OMNI), texture type, respectively. In this, ENT and HOMO represent entrophy and homogeneity, respectively. As shown in Fig. 3, two results is only different direction with other same parameters. The interpretation of texture image is somewhat complicated, and direction dependency for texture measure is still problematic. Therefore, the omni-direction is helpful to summarize texture measures.
Fig. 3. (A) (8, GLCM, 5*5, E-W, ENT), (B) (8, GLCM, 5*5, NW, ENT).
Fig. 4 shows application cases related to GLCM and GLDV. Fig. 4(A) and (B) are from same quantization level, kernel size, direction, and texture type, except application scheme. In the urban remote sensing containing urban feature characterization or classification, both texture images show significant implication: detection of shadow zone, classification of building types, and recognition of pavement condition in the micro-scale.
This implication dealing with texture measures is not a final one, and some further works are needed: applicability of GLCM and GLDV to urban remote sensing, and selection guide of proper type among texture measures to characterize complicated urban features.
Fig. 4. (A) (16, GLCM, 7*7, OMNI, HOMO), (B) (16, GLDV, 7*7, OMNI, HOMO).
With these preliminary works concerned texture measures, texture images are also utilized to data fusion. Fig. 5 represents two cases of texture image fusion, and applied data is covering urban area containing complex urban features shown in Fig. 1.
This approach is based on HIS (Hue-Intensity-Saturation), one of popular or basic data fusion schems. Input images in Fig. 5(A) are GLCM-based entrophy images by same quantization level, and omni-direction, but each kernel size differ from each other as 11*11, 7*7 and 3*3. While, input images in Fig. 5(B) are different quantization level, application scheme, type, with same kernel size and direction. Though this result is not studied in detailed, two cases of (A) and (B) reveals common distinguishable features, in the visual interpretation.
Fig. 5. Fused cases: (A) HIS = (16, GLCM, 11*11, OMNI, ENT), (16, GLCM, 3*3, OMNI, ENT), (16, GLCM, 7*7, OMNI, ENT), (B) HIS = (2, GLCM, 5*5, OMNI, HOMO), (16, GLDV, 5*5, OMNI, ENT), (8, GLCM, 5*5, OMNI, Energy).
5. Concluding Remarks and Further works
Most previous studies for second order texture analysis have been toward the improvement of classification accuracy, with supervised or un-supervised classification methods, showing high accuracy. Scope of this study is somewhat different from previous works. An application program for texture measures based on GLCM and GLDV is newly implemented in this study. By using this program, GLCM or GLCV-based texture images by different quantization level, kernel size, and texture type are created with the high-resolution satellite image covering urban area; in this study, omni-direction texture measures are regarded as useful one to reduce directional dependency which causes difficulty in texture image interpretation. In application of urban feature characterization to texture measures, texture images is helpful to detect shadow zone, classify building types, and recognize road pavement condition in the micro-scale, though they are not fully analyzed in this study. As for the data fusion perspective, as well as these advantageous aspects, selection reference for applicable type among multiple texture measures or application schemes, to properly reveal the complicated urban features, are necessary with actual examples.
- Al-Janobi, A., 2001, Performing evaluation of cross-diagonal texture matrix method of texture analysis, Pattern Recognition, 34, pp.171-180.
- Bharati, M. H, Liu, J. J., and MacGregor, J. F., 2004, Image texture analysis: methods and comparisons, Chemometrics and Intelligent Systems, 72, pp. 57-71.
- Clausi, D. A. and Zhao, Y., 2003. Grey level co-occurrence integrated algorithm (GLCIA): a superior computational method to rapidly determine co-occurrence probability texture features, Computers & Geosciences, 29, pp. 837-850.
- Demin X., 2002, Remote Sense and GIS-based Evacuation Analysis, ORNL presentation material, NCRST 2002.
- Dulyakarn, P, Rangsanseri, Y., and Thitimajshima, P., 2000, Comparison of Two Features for Multispectral Imagery Analysis, Proceeding of Asian Conference of Remote Sensing.
- Franklin, S. E., Wulder, M. A., and Gerylo, G. R., 2001, Texture analysis of IKONOS panchromatic data for Douglas-fir forest age class separability in British Columbia, Int. Jour. of Remote Sensing, 22, pp. 2676-2632.
- Hall-Beyer, M., 2004, GLCM Texture: A Tutorial v.2.7.1, on-line document, .
- Haralick, R. M., Shanmugam, K. and Dinstein, I., 1973, Textural features for image classification, IEEE Trans. Sys. Man. Cybern., SMC-3, pp. 610-621.
- Herold, M. H, Liu, X., and Clake, K. C., 2003, Spatial Metrics and Image Texture for Mapping Urban Land Use, PE&RS, 69, pp. 991-1001.
- Maillard, P., 2003, Comparing Texture Analysis Methods through Classification, PE&RS, 69(4), pp. 357-367.
- Kiema, J.B.K, 2002, Texture analysis and data fusion in the extraction of topographic object from satellite imagery, Int. Jour. of Remote Sensing, 23(4), pp. 767-776.
- Parker, J. R., 1997, Algorithms for Image Processing and Computer Vision, John Wiley & Sons, pp. 150-175.
- Smith, A. M. S., Wooster, M. J., Powell, A. K., and Usher, D., 2002, Texture based feature extraction: application to burn scar detection on Earth observation satellite sensor imagery, Int. Jour. Remote Sensing, 23, pp. 1733-1739.
- Wang, X. and Hanson, A. R., 2001, Surface texture and microstructure extraction from multiple aerial images, Computer Vision and Image Understanding, 83, pp. 1-37.
- Zhang, Y, 1999, Optimisation of building detection in satellite images by combining multispectral classification and texture filtering, ISPRS Journal of Photogrammetry & Remote Sensing 54, pp. 50-60.