Image Registration using neural networks as a shape discriminator

Image Registration (IR) is a process of transforming images in a dataset into one common coordinate system. The original method, developed by Mangan and Shiflet for metallurgical microstructures, relied on hardness indentations to align serial images in the dataset. However, as the sample is continuously sectioned, the marks and the microstructure change and an appropriate 'hands-on' registration method is required to locate the marks for IR. Typical feature and area based algorithms fail to perform adequately when marks are scaled or rotated. A Neural Network (NN) can be trained to discriminate shapes of interest which can then be used to assist and automate the IR process. ANN that can successfully isolate indentation squares that are rotated, scaled and located among other microstructural features will be presented. That information is then used to automatically register a set of synthetic and optical microscope images.