Content area
Abstract
The development of high resolution CCD-cameras has resulted in image data reduction becoming an important part of industrial image processing. At the same time new multimedia techniques have been accompanied by the development of new data reduction algorithms which are able to achieve good compression ratios. These algorithms are however optimised for human vision and can not achieve acceptable results in many cases for industrial image processing applications.
This thesis investigates the possibility of focusing on important regions within a grey-scale image in order to achieve good data reduction rates ($>$1:10) for industrial image processing tasks. The main aspect of all these methods is to vary the resolution or reduction rate depending on local image content. The common aim of the methods studied is to preserve the object contour and edge information while minimising the total image data. The technique is to use intelligent image data reduction methods optimised for industrial image processing tasks.
Different foveal algorithms are described and investigated within this thesis. These foval algorithms try to simulate the functionality of a human eye in order to achieve good data reduction rates by varying the resolution within one frame.
A detailed description and comparison is presented of common methods for reducing image data by focusing on regions of interest (FORI). A novel approach requiring no a priori knowledge is presented that uses DCT coefficients for the detection of important image areas. In order to show the usability of this new FORI method, this thesis presents a new method for context-sensitive image data reduction.
The experiments presented in this thesis provide evidence that common compression methods achieve insufficient data reduction when used for measurement tasks in industrial image processing. It is shown that the new FORI method can reduce the errors at object boundaries and edges while achieving data reduction ratios comparable with common multimedia data reduction algorithms.
This thesis demonstrates the suitability of the new approach of combining region focusing algorithms and industrial image processing requirements into a novel context-sensitive data reduction algorithm.
