Model Based System Development/Move or Clean UP/Image Processing

Image Processing (805A-10)

Staff

• Theo Schouten (project leader)

Start/End Date

This project started 1-1-96. Currently no end date is foreseen.

Description

The aim of this project is to develop methods for obtaining useful information from image-like sensor data. One aspect of this is `recognition': determine the presence of features or objects, a task in which the human cognitive systems still excels. Another aspect is `measurement': determine quantitative properties of (recognized) features and objects. There are many technical application area's for these image processing aspects.

Currently the focus of this project is changed from satellite image processing to processing of images and video for robotics systems. A small robot lab is now set up using robots build from Lego parts and webcam like video sources. This lab will also provide testbeds for other research activities within the RTHS programme and be used for students projects and in teaching.

Previously the focus was on crop area estimation from satellite images, which involves not only the recognition of agricultural fields and crops grown on them, but also the accurate measurement of the sizes of those fields. The limited spectral and spatial resolution of the satellite sensors complicates this task. The amount of data is very large and parallel methods must be developed to achieve the task in a reasonable amount of time. The methods must also be adaptive as the characteristics of crops change with time and location due to natural factors. Therefore, neural network methods were investigated because of their learning capabilities.

Results from the robot lab

The basic infrastructure was setup and has been used for student projects and in teaching.

Results on satellite image processing

The dissertation of M. klein Gebbinck "Decomposition of mixed pixels in remote sensing images to improve the area estimation of agricultural fields" was finished in 1998.

Research concentrated on decomposition of mixed pixels using neural networks methods. A new image with a ground truth data set was acquired and a new method was developed to purify and enlarge ground truth data sets using the corresponding images. A large number of neural network training runs were done on the PowerXplorer parallel machine to determine whether the neural network is also able to handle a large number of ground classes. Results on this research were presented at the Florence conference in 1999 under "Fuzzy classification of pixels using neural networks".

From september to december 1999 prof. Zhenkai Liu from the USTC, Hefei, China visted us. Research now concentrated on the development of hierarchical neural networks. To obtain sufficient computing speed the parallel neural network simulator was transformed to a serial version running on a SUN workstation. Formats for input, intermediate and output files were kept the same, in order to use both machines for the samen experiment. A first result was presented at the Barcelona conference in 2000 under "Hierarchical neural networks for pixel classification".

Planning

Continuing the buildup of the robot lab.