Machine vision techniques for inspection of dry-fibre composite preforms in the aerospace industry

Abstract

This thesis presents the results of a three year investigation into machine vision techniques for in-process automated inspection of dry-fibre composite preforms. Efficient texture analysis based techniques have been developed, tested, and implemented in a prototype robotic assembly cell. Industrial constraints have been considered in the development of all the algorithms described. A single channel texture analysis model is described which can successfully segment images containing only a few textures. The model is based on convolution of the image with small kernels optimised for the task, and is elegant in the sense that it is computationally simple and easily realisable in low cost hardware. A new convolution kernel optimisation algorithm is described. It is demonstrated that convolution kernels can also be optimised to perform as edge operators in simple textured images. A novel boundary refinement algorithm is described which reduces the inspection errors inherent in texture based boundary estimates. The algorithm takes the form of a local search, using the texture estimate as a guiding template, and selects edge points by maximising a merit function. Optimum parameters for the merit function are obtained using multiple training images in conjunction with simple function optimisation algorithms.