Multi-Directional Projection Transformations for Machine Learning based Object Matching

In case of object matching on low-quality images, projection-based similarity measurement is an often used approach. In recent advances, deep convolutional neural networks are used to extract the features from the input data. In this paper, an analysis is done to compare end-to-end methods with projection-based preprocessing methods, including the Radon transform and the Trace transform. The paper explains the method of convolutional neural architecture generation based on input sizes, followed by the parallel processing on a cluster of workstations. To show the applicability of multi-directional projections as a preprocessing method on small storage units, model memory cost and prediction performance are both measured and optimal results are selected accordingly.