Hyperspectral Image Segmentation using Active Contour and Graph Cut Susi Huaman De la Vega, MS Graduate Student UPRM,
[email protected] Dr. Vidya Manian, Assistant Professor UPRM,
[email protected] Laboratory for Applied Remote Sensing and Image Processing University of Puerto Rico at Mayagüez, P. O. Box 9048, Mayagüez, Puerto Rico 00681-9048
ABSTRACT In this research, an algorithm that combines active contours and graph cut approaches for hyperspectral image segmentation is developing. Active contours have been widely used as attractive image segmentation methods because they produce sub-regions with continuous boundaries. On the other hand graph cuts have emerged as a powerful optimization technique for minimizing energy functions and avoid the problems of local minima inherent in other approaches.
Figure 5: (a) Region selected from HYDICE Terrain sample scene. This image contains 60 lines, 83 samples and 210 bands. (b) Initial contour generated based on the spectral signature. (c) Result of the segmentation in the band 93.
The combination of two models can show robust hyperspectral image segmentation capability, because it has the ability to jump over local minima and provide a more global result and graph cuts guarantee continuity and lead to smooth contours free of self-crossing and uneven spacing problems. Algorithm validating and comparisons are done with real and synthetic hyperspectral images. This algorithm can be applied in many fields and it will represent an important advantage to the segmentation field.
Figure 6: (a) Region selected from HYDICE Urban sample scene. This image contains 29 lines, 44 samples and 210 bands. (b) Initial contour generated based on the spectral signature. (c) Result of the segmentation in the band 1.
TECHNICAL APPROACH Active Contour Active contour lies on the idea of deforming the initial curve to the boundary of objects under some constraints from the image using techniques of curve evolution[Ning03]. There are two difficulties in the design and implementation of active contour models. In the first place, the initial contour must be close to the true boundary. Second, active contours have difficulty in processing into boundary concavities [Li09]. Graph Cut
Figure 2: Block diagram of the Algorithm.
Figure 7: (a) Region selected from HYDICE Forest sample scene. This image contains 120 lines, 152 samples and 210 bands. (b) Initial contour generated based on the spectral signature. (c) Result of the segmentation in the band 84. (c) Result of the segmentation in the band 84. (d) Final result with the correction points given by the user based in [Ning03].
PRELIMINARY RESULTS
FUTURE PLANS
For our preliminaries results we used different hyperspectral images, taken by different sensors, to obtain the initial contour we use the spectral angle metric with different values for threshold; we can observed that our approach have acceptable results:
1. Construct adjacency matrix for all the bands in the image and apply the graph cut method. 2. Segment more than one target in the image. 3. Try to define initial contour in an unsupervised manner.
REFERENCES
Recently, a fast energy minimization techniques based on graph cuts have emerged. These techniques can be applied to a restricted class of energy functions of discrete variables. An advantage of these methods is that in certain cases they can produce a global minimum of the energy or in other cases a local minimum with some strong properties [Ning03].
METHODOLOGY Figure 3: (a) Region selected from Fake Leaves hyperspectral image captured by SOC-700. This image contains 360 lines, 360 samples and 120 bands. (b) Initial contour generated based on the spectral signature. (c) Result of the segmentation in the band 25.
[Boykov06] Yuri Boykov and Gareth Funka-Lea, “Graph Cuts and Efficient N-D Image Segmentation,” Int. J. Comput. Vision 70, no. 2 (2006): 109-131. [Cheolha05] Cheolha, P. L., “Robust Image Segmentation using Active Contours: Level Set Approaches”, thesis (PHD) In North Carolina State University (USA, 2005). [Li09] Li G., Sun X., Zheng Y., Zhou X., “Geometric Active Contours Withoutre-Initialization for Image Segmentation”, available at Science Direct, Pattern Recognition, In Press, Corrected Proof, (Jan. 2009). [Ning03] Ning Xu, R. Bansal, and N. Ahuja, “Object segmentation using graph cuts based active contours,” in Computer Vision and Pattern Recognition, 2003. Proceedings. 2003 IEEE Computer Society Conference on, vol. 2, 2003, II-46-53 vol.2. [Rivera04] Rivera, F., “Segmentation of Underwater Multispectral Images with Applications in the Study of Coral Reefs”, Thesis (MS) in Electrical Engineering University of Puerto Rico - Mayaguez Campus (Puerto Rico, May 2004).
STRATEGIC RESEARCH PLAN This work will be useful for CenSSIS Researchers and Students from R2, S3, and S4 who make use of multi and hyperspectral images and will result in technology transfer to the industry in the form of tools and methodologies for spectral image processing.
Figure 1: General block diagram for the Hyperspectral Image Segmentation Algorithm.
Figure 4: (a) Region selected from Washington D.C. Mall hyperspectral image. This image contains 320 lines, 250 samples and 191 bands collected from 0.4 to 2.4 µm of the visible and infrared spectrum. (b) Initial contour generated based on the spectral signature. (c) Result of the segmentation in the band 181.
This work was supported by Gordon-CenSSIS, the Bernard M. Gordon Center for Subsurface Sensing and Imaging Systems, under the Engineering Research Centers Program of the National Science Foundation (Award Number EEC-9986821).