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

Search

Browse

My Space

U-NET Neural network analysis and implementation using reconfigurable logic

Skoufis Charalampos

Full record


URI: http://purl.tuc.gr/dl/dias/4F35DDD9-7805-43D0-B8FD-E366D4628B0D
Year 2021
Type of Item Diploma Work
License
Details
Bibliographic Citation Charalampos Skoufis, "U-NET Neural network analysis and implementation using reconfigurable logic", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2021 https://doi.org/10.26233/heallink.tuc.88479
Appears in Collections

Summary

In recent years, neural networks are increasingly the primary tool for image analysis, providing exceptional accuracy vs. human perception. In the field of biomedicine, in particular, the misdiagnosis of magnetic resonance imaging or computed tomography (MRI / CT) scans is a significant problem in preventing and treating various health problems which are impossible to detect by the human eye. In the field of terrain pattern recognition performed by power-limited mini-satellites, a more efficient approach for both architecture and hardware equipment is required. A recent U-shaped architecture offers impressive results and methods for detecting patterns and anomalies using semantic image segmentation. This thesis work is based on this U-NET architecture and aims to analyze, model, and build the network on multiple programming levels of abstraction, including hardware. At present, there exist more mature architectures such as Convolutional Neural Networks (CNN) that have substantial support toolsets.On the other hand, U-NET architecture does not have a great level of support tools; this work will try to address this issue. The main structure and learning process (training) of this neural network will also be presented in detail, along with all the additional tools to assist this process. The code pack starts with a user-friendly Python language, where user-customizable functions and training techniques will be introduced.The Python language level is intended mostly to aid the learning process. One step further, researchers can proceed by utilizing the C language, where the prediction step has been constructed to be further analyzed and eventually reach a specific application platform. Finally, three building blocks of this network have been implemented on Field Programmable Gate Array (FPGA) and Graphics Processor Unit (GPU) platforms (on par with the entire NN), offering the acceleration of specific processes with substantial energy savings for the computation. Last, but not least, the ecosystem developed in this thesis was not available until now - with its use more researchers can efficiently employ U-NETs.

Available Files

Services

Statistics