Institutional Repository [SANDBOX]
Technical University of Crete
EN  |  EL

Search

Browse

My Space

Study of a rotationally invariant hardware implementable convolutional neural network using CORDIC arithmetic

Michail Sotirios

Full record


URI: http://purl.tuc.gr/dl/dias/D10A98E6-AFA7-499F-BB74-CC4B5A02BE95
Year 2024
Type of Item Diploma Work
License
Details
Bibliographic Citation Sotirios Michail, "Study of a rotationally invariant hardware implementable convolutional neural network using CORDIC arithmetic", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2024 https://doi.org/10.26233/heallink.tuc.98959
Appears in Collections

Summary

Introduced in this thesis is an approach in enhancing the rotational invariance of Convolutional Neural Networks (CNNs), through integrating the novel Log-CORDIC algorithm for image pre-processing. This image pre-processing algorithm presents an advantage over existing cartesian-to-polar transform algorithms for images, through the computational advantages of the Coordinate Rotation Digital Computer (CORDIC) algorithm. The results of the novel algorithm are studied and compared with existing transform methods, along with its efficiency improvements, and its ability to enhance rotational invariance in a CNN is ascertained by integrating it into the pipeline of a customized SqueezeNet neural network. Focusing on the CIFAR-10 and MNIST datasets, experiments with this customized SqueezeNet neural network demonstrate an improvement in classification accuracy for images with varied orientations.

Available Files

Services

Statistics