Image processing for amorphous foodstuffs

This thesis focuses on the problems of using image processing for the task of identifying amorphous objects in the food industry. To aid this investigation a system has been constructed to identify amorphous overlapping objects where specific elements of the image may be partially or completely obscured. Furthermore the elements may vary greatly in shape and basic image properties. The recognition system has a layered architecture. The low level layer consists of pattern classifiers using colour and texture. At the higher layers a rule based system uses domain specific knowledge to uniquely label the image elements that, due to the high variability of their basic image properties cannot be identified by the lower layers. The system is applied to the specific task of recognising certain key elements within the viscera of a bovine carcass varying in weight from 200-800kg (and associated breed, sex and age differences) previously automatically ejected onto a conveyor from the ventral cavity of the animal. The performance achieved is similar to that of a human working with the same images. The thesis argues that there is an intrinsic use of knowledge with image processing solutions to achieve this level of performance and at every stage of processing knowledge must be incorporated into the system.