Using texture to tackle the problem of scale in land-cover classification
3.4 Segmentation evaluation
pared to texture. If a boundary has both an intensity and texture boundary, the crest of the gradient magnitude image will be located at the intensity boundary. The watershed algorithm will therefore locate boundaries using an intensity boundary as opposed to a texture boundary where possible and minimize spatial uncertainty. This effect is illustrated in Fig. 9.
ries between tree tops, localization is not so accurate. Some over- and un- der-segmentation is also evident in highly textured regions.
(a) (b)
(c) (d) Fig. 11 Segmentation results obtained from the segmentation procedure described.
Primitive-object boundaries are represented by the colour white. Ordnance Survey Crown Copyright. All rights reserved
We performed quantitative performance evaluation in an unsupervised manner, which does not require ground-truth, using an unsupervised per- formance metric know as the SU metric (in forthcoming publication).
Evaluation of remotely sensed data segmentation has been performed in such an unsupervised manner in the past (Pal and Mitra 2002; Mitra et al.
2004). The SU metric calculates a ratio of the individual segments texture and intensity feature separation to cohesion, with higher values indicating better segmentation performance. This metric has been shown to be an ac- curate performance metric having a high correlation with an accurate su- pervised metric on a synthetic dataset, containing noisy uniform and tex- tured regions, where accurate ground-truth is known. This correlation is significantly greater than previous attempts to establish a relationship be- tween supervised and unsupervised metrics on the same data (in forthcom- ing publication). We evaluated our segmentation algorithm against the marker-controlled watershed transform applied to gradients extracted using
the Canny gradient detector (Canny 1986). These gradients were calcu- lated by smoothing with a Gaussian of sigma 1.5 followed by application of the Sobel gradient operator. A set of 60 images of size 256x256 pixels was used for evaluation. This set was divided into 20 training images to optimize the segmentation scale parameter H for each algorithm and 40 test images. Our proposed segmentation algorithm achieved an average SU metric value of 0.43 on the test data set. This result outperformed the marker-controlled watershed transform applied to the Canny gradients which accomplished an average SU metric value of 0.38 on the same data.
4 Conclusions and Future Work
Generating accurate land-use classification in urban areas from remotely sensed imagery is a challenging problem. The ultimate goal of OBIA is to model the HVS, which can perform this task quite easily. We argue that some previous implementations of the OBIA paradigm are based on an in- accurate conceptual model of the HVS. We therefore proposed a new con- ceptual model which we feel overcomes this weakness and will hopefully lead to more accurate land-use classification systems in the future.
We attempted to implement the first step in this conceptual model by performing segmentation where each area of uniform visual properties is segmented correctly. To achieve this we proposed a segmentation algo- rithm which involves the computation and fusion of a novel complemen- tary texture and intensity gradient set. The segmentation results achieved are visually very accurate, although some over- and under-segmentation is evident. Using an unsupervised performance metric we showed that the proposed algorithm quantitatively outperforms the marker-controlled wa- tershed transform applied to gradients extracted using the popular Canny gradient operator.
The slight over- and under-segmentation evident in the results may be due to gradient images being based on a local measure of difference which does not encode information regarding the difference between the interior of regions (O'Callaghan and Bull 2005). A potential solution may be post- processing of the segmentation in which regions are merged based on their interior region properties. Some primitive-object boundaries suffer from poor localization. This is due to texture gradients only being localized to the spatial scale of feature extraction. Future work will attempt to localize these gradients to a finer spatial scale.
References
Baatz M, Schape A (2000), Multiresolution segmentation - an optimization ap- proach for high quality multi-scale image segmentation, J. et al. (eds.): Ange- wandte Geographische Informationsverarbeitung XIII, Wichmann, Heidel- berg, pp 12-23
Benz UC, Hofmann P, Willhauck G, Lingenfelder I, Heynen M (2004), Multi- resolution, object-orientated fuzzy analysis of remote sensing data or GIS- ready information, ISPRS Journal of Photogrammetry & Remote Sensing, vol 58, pp 239-258
Biederman I (1987), Recognition-by-components: a theory of human image un- derstanding, Psychol Rev, vol 94, pp 115-147
Black M, Sapiro G (1999) Edges as outliers: anisotropic smoothing using local image statistics. International Conference on Scale-Space Theories in Com- puter Vision, Corfu, Greece, pp 259-270
Black MJ, Sapiro G, Marimont DH, Heeger D (1998), Robust anisotropic diffu- sion, IEEE Trans on Image Process, vol 7, no 3, pp 421-432
Blaschke T (2003) Object-based contextual image classification built on image segmentation. IEEE Workshop on Advances in Techniques for Analysis of Remotely Sensed Data, pp 113-119
Canny JF (1986), A computational approach to edge detection, IEEE Trans Pat- tern Anal and Mach Intell, vol 8, no 6, pp 679-698
Chaji N, Ghassemian H (2006), Texture-gradient-based contour detection, EURASIP Journal on Applied Signal Processing, vol 2006, pp 1-8
Clausi DA, Jernigan ME (2000), Designing Gabor filters for optimal texture sepa- rability, Pattern Recogn, vol 33, no 11, pp 1835-1849
Corcoran P, Winstanley A (2007) Removing the texture feature response to object boundaries. International Conference on Computer Vision Theory and Appli- cations, Barcelona, pp 363-368
Deng H, Liu J (2003), Development of anisotropic diffusion to segment texture images, J Electron Imag, vol 12, no 2, pp 307-316.
Gonzalez R, woods R, Eddins S (2003) Digital image processing using matlab, Prentice Hall
Julesz B (1983), Textons, the fundamental elements in preattentive vision and per- ception of textures, Bell System Technical Journal, vol 62, pp 1619-1645 Malik J, Belongie S, Leung T,Shi J (2001), Contour and texture analysis for image
segmentation, Int J of Comput Vis, vol 43, no 1, pp 7-27
Mitra P, Shankar BU, Pal SK (2004), Segmentation of multispectral remote sens- ing images using active support vector machines, Pattern Recogn Lett, vol 25, no 2004, pp 1067-1074
O'Callaghan RJ, Bull DR (2005), Combined morphological-spectral unsupervised image segmentation, IEEE Trans Image Process, vol 14, no 1, pp 49-62 Pal SK, Mitra P (2002), Multispectral image segmentation using the rough-set-
initialized EM algorithm, IEEE Trans Geosci Rem Sens, vol 40, no 11, pp 2495-2501
Palmer SE (1999) Vision science: photons to phenomenology, The MIT Press.
Palmer SE, Rock I (1994), Rethinking perceptual organization: The role of uni- form connectedness, Psychonomic Bull Rev, vol 1, no 1, pp 29-55
Perona P, Malik J (1990), Scale-space and edge detection using anisotropic diffu- sion, IEEE Trans Pattern Anal Mach Intell, vol 12, no 7, pp 629-639
Pylyshyn ZW (1999), Is vision continuous with cognition? The case for cognition impenetrability of vision perception, Behav Brain Sci, vol 22, no 3, pp 341- 365
Shao J, Forstner W (1994) Gabor wavelets for texture edge extraction. ISPRS Commission III Symposium on Spatial Information from Digital Photogram- metry and Computer Vision, Munich, Germany, pp 745-752
Soille P (2002) Morphological image analysis: principles and applications, Springer-Verlag Berlin and Heidelberg GmbH & Co. K
Treisman A, Gelade G (1980), A feature integration theory of attention, Cognit Psychol, vol 12, pp 97-136
Vecera SP, Farah MJ (1997), Is visual image segmentation a bottom-up or an in- teractive process?, Percept Psychophys, vol 59, no 8, pp 1280-1296
Watt RJ (1995) Some speculations on the role of texture processing in visual per- ception. In: T. V. Papathomas, C. Chubb, A. Gorea and E. Kowler (eds) Early Vision and Beyond, The MIT Press, pp 59-67