AbstractThe aim of this research work was to obtain a system that classifies texture. This so called Texture Classification System is not a system for one special task or group of tasks. It is a general approach that shows a way towards real artificial vision.
Finding ways to enable computerised systems to visually recognise its surroundings is of increasing importance for the industry and society at large. To reach this goal not only objects but less well describable texture has to be identified within an image.
To achieve this aim a number of objectives had to be met. At first a review of how natural vision works was done to better understand the complexity of visual systems. This is followed by a more detailed definition of what texture is. Next a review of image processing techniques, of statistical methods and of soft-computing methods was made to identify those that can be used or improved for the Texture Classification System.
A major objective was to create the structure of the Texture Classification System. The design presented in this work is the framework for a multitude of modules arranged in groups and layers with multiple feedback and optimisation possibilities.
The main achievement is a system for texture classification for which natural vision was used as a " blue-print". A more detailed definition of what texture is was made and a new texture library was started. The close review of image processing techniques provided a variety of applicable methods, as did the review and enhancement of statistical methods. Some of those methods were improved or used in a new way. Neural networks and fuzzy clustering were applied for classification, while genetic algorithms provide a means for self optimisation.
The concepts and methods have been used for a number of projects next to texture classification itself. This work presents applications for fault detection in glass container manufacturing, quality control of veneer, positioning control of steel blocks in a rotation oven, and measurement of hair gloss.
With the Texture Classification System a new, holistic approach for complex image processing and artificial vision tasks is being contributed. It uses a modular combination of statistics, image processing and soft-computing methods, easily adaptable to new tasks, includes new ideas for high order statistics, and incorporates self optimisation to achieve lean sub-systems. The system allows multiple feedbacks and includes a border detection
routine. The new texture library provides images for future work of researchers.
Still a lot of work has to be done in the future to achieve an artificial vision system that is comparable to the human's visual capabilities. This is mainly due to the fact of missing computational resources. At least another decade of hardware development is needed to reach this goal. During this time more, better or even novel methods will be added to the Texture Classification System to improve its universal capabilities.
|Date of Award
|Robert Williams (Supervisor)