Innovative Machine Vision Technique for 2D/3D Complex and Irregular Surfaces Modelling

This study propose and demonstrates a novel technique incorporating multilayer perceptron (MLP) neural networks for feature extraction with Photometric stereo based image capture techniques for the analysis of complex and irregular 2D profiles and 3D surfaces. In order to develop the method and to ensure that it is capable of modelling non-axisymmetric and complex 2D/3D profiles, the network was initially trained and tested on 2D profiles, and subsequently using objects consisting of between 1 and 4 hemispherical 3D forms. To test the capability of the proposed model, random noise was added to 2D profiles. 3D objects were coated with various degrees of coarsenesses (ranging from low-high). The gradient of each surface normal was quantified in terms of the slant and tilt angles of the vector about the x and y axis respectively. The slant and tilt angles were obtained from the bump maps and these data were subsequently employed for training of a NN that had x and y as inputs and slant and tilt angles as outputs. The network employed had the following architecture: MLP and a Levenberg-Marquardt algorithm (LMA) for training the network for 12,000 epochs. At each point on the surface the network was consulted to predict slant and tilt and the actual slant and tilt was subtracted, giving a measure of surface irregularity. The network was able to model the underlying asymmetrical geometry with an accuracy regression analysis R-value of 0.93 for a single 3D hemispheres and 0.90 for four adjacent 3D non-axisymmetric hemispheres.