In this paper we propose different FPGA implementations of a Convolutional Neural Network where the use case is to detect road cracks via images. The work is based on the network proposed by the current state of the art in the field. We deploy the network using the MATLAB Deep Learning HDL toolbox on all available AMD-Xilinx platforms, including different data types. In particular, we use the ZC706 and ZCU102 development boards. In order to infer the CNN, we apply a single precision and an 8-bit integer data type quantization. The implementation results show that the detection accuracy of 99.6% is the same of the state of the art, even though the network is quantized. We also obtain a speed-up of the CNN reaching up to 313.2 Frames Per Second while requiring only 45.85 mJ to process one frame. The proposed implementations are therefore a viable solution for a fast and low-power crack detection system.
Canese, L., Cardarilli, G., Di Nunzio, L., Fazzolari, R., Re, M., Spano, S. (2023). FPGA-Based Road Crack Detection Using Deep Learning. In International Conference on System-Integrated Intelligence: SYSINT 2022: Advances in System-Integrated Intelligence (pp.65-73). GEWERBESTRASSE 11, CHAM, CH-6330, SWITZERLAND : SPRINGER INTERNATIONAL PUBLISHING AG [10.1007/978-3-031-16281-7_7].
FPGA-Based Road Crack Detection Using Deep Learning
Cardarilli, GC;Di Nunzio, L;Fazzolari, R;Re, M;Spano, S
2023-01-01
Abstract
In this paper we propose different FPGA implementations of a Convolutional Neural Network where the use case is to detect road cracks via images. The work is based on the network proposed by the current state of the art in the field. We deploy the network using the MATLAB Deep Learning HDL toolbox on all available AMD-Xilinx platforms, including different data types. In particular, we use the ZC706 and ZCU102 development boards. In order to infer the CNN, we apply a single precision and an 8-bit integer data type quantization. The implementation results show that the detection accuracy of 99.6% is the same of the state of the art, even though the network is quantized. We also obtain a speed-up of the CNN reaching up to 313.2 Frames Per Second while requiring only 45.85 mJ to process one frame. The proposed implementations are therefore a viable solution for a fast and low-power crack detection system.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.