Print
Univis
Search
 
FAU-Logo
Techn. Fakultät Website deprecated and outdated. Click here for the new site. FAU-Logo

Dipl.-Inf. Johannes Jordan

Alumnus of the Pattern Recognition Lab of the Friedrich-Alexander-Universität Erlangen-Nürnberg

Hyperspectral cameras are a core technology in remote sensing. Recently, multispectral and hyperspectral sensors became increasingly applicable to other fields, e.g. cultural heritage and medical diagnosis. A camera that can see more information than the human eye is fascinating. My vision is that we can access this unseen information easily and intuitively with the help of novel processing and visualization systems.

Multispectral Image Segmentation

true-color display
pixel labels
spectral distribution

A vital analysis tool within our framework is segmentation. With a good segmentation, the user can easily find specific regions in the image, reveal charasteristics of these regions, and contrast them to others based on their content. As an example, in the picture with fake and real peppers to the right, their difference is easily revealed just by automatically segmenting both and looking at their spectral distribution.

To bring segmentation to our framework, we investigate clustering and segmentation algorithms that are popular in computer vision and find adaptations for the multispectral domain.

 

Supervised Segmentation
true-color display
seed point input
edge weights in x-direction
segmentation result

In an interactive analysis framework, input from the user is an important prior to the segmentation tasks. As the user explores the multispectral data step-by-step, she may want to compare the spectra of specific objects in the scene or examine reflectance properties of a certain area in detail. Such a segmentation replaces tedious manual labeling of this area, but was previously unavailable for multispectral or hyperspectral images.

In recent years, graph-based methods have had a significant impact on image segmentation. They are especially noteworthy for supervised segmentation, where the user provides task-specific foreground and background seeds. In our work we adapted the graph-cut and power watershed algorithms to the multispectral domain. For this, we define new edge weights in the graph based on spectral dissimilarity measures. We also introduce a new data-driven measure using the self-organizing map.

 

Publications:

Jordan, Johannes; Angelopoulou, Elli. Supervised Multispectral Image Segmentation With Power Watersheds. 19th IEEE International Conference on Image Processing (ICIP), Orlando, FL, 30.09.2012, pp. 1585-1588, 2012

Unsupervised Segmentation
true-color display
clustering via mean shift on image data
clustering via mean shift on spectral gradient

In our framework, the user explores materials and reflectance effects in a scene captured by a multispectral camera without the use of application-dependent dimensionality reduction or pixel classification. A viable tool for such analysis is clastering of the data. In particular, the mean shift algorithm is a well-understood and popular clustering method that is in theory applicable to high-dimensional data, such as multispectral pixel vectors.

In this project, we examine recent variants of the mean shift method and develop new algorithms that reduce the computational complexity of the method. Our goal is to obtain a high-quality global segmentation in a fraction of time needed to cluster a multispectral image with current methods.

 

Publications:

Jordan, Johannes; Angelopoulou, Elli. Mean-shift Clustering For Interactive Multispectral Image Analysis. 20th IEEE International Conference on Image Processing (ICIP), to appear