Image recognition performed on handwritten letters using the windowed scattering transform
Abstract
The windowed scattering transform is an operator that is invariant to small translations,deformations and rotations. The transform can be used in conjunctionwith a classification algorithm to perform image recognition. This thesis consistsof one theoretical part and one numerical part. In the theoretical part the underlyingtheory of the windowed scattering transform, namely Fourier analysis andwavelets, is briefly introduced. Then, the construction of the windowed scatteringtransform and its numerical approximation is explained in detail. The numericalpart consists of examples showcasing the properties of the transform, andthe transform applied in image recognition on a dataset of handwritten letters.An error rate of 10.2% was achieved, using the k-nearest neighbors algorithm forclassification. The error rate is high compared to other more sophisticated imagerecognition procedures. Most of the errors stem from inaccurate classification onclasses with few samples, and from incorrect classifications on letters that are similarin shape. Some suggestions are given on how the error rates could be improvedin further work.