Institutional Repository
Technical University of Crete
Technical University of Crete
EN
|
EL
| URI: | http://purl.tuc.gr/dl/dias/2FF416B3-AC4F-4F68-BFD5-FD47687B252B | ||
| Year | 2024 | ||
| Type of Item | Diploma Work | ||
| License |
|
||
| Bibliographic Citation | Alexandros-Ioannis Sifakis, "FPGA-Based embedded system to detect cracks in harbor structures with the Use of convolutional neural network", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2024 https://doi.org/10.26233/heallink.tuc.101257 | ||
| Appears in Collections |
Convolutional Neural Networks are highly effective in a wide range of applications, particularly in the field of computer vision, where they perform very well at recognizing patterns and objects. One key application of CNNs is to detect cracks in harbor structures. In this thesis, an FPGA-based accelerator intellectual property core for this application was developed, based on a pre-existing Convolutional Neural Network. The accelerator will be integrated with a RISC-V computing core for real-time, on-site crack detection. The large memory footprint of the CNN weights (> 130MB) did not allow for all weights be be internally stored in the FPGA's 71MB space. To address this, extensive experimentation with the K-means algorithm was performed in order to effectively compress the floating point weights so that they would fit insid the FPGAs's memory. This was achieved, resulting in a 4X compression, while maintaining at least 95% accuracy vs. the reference CNN. The CNN, a UNET of no less than 24 stages, was designed, synthesized and functionally verified with the Xilinx Vitis HLS CAD tool suite. Functional verification was completed successfully, using the same dataset that was used to train and evaluate the model. The IP core can be integrated as-is with the RISC V processor on the Alveo U55C system, however, several changes have been proposed in this thesis to improve performance.
Copyright © DIAS 2013
