CNN implementation on FPGA with automated environments for computer vision

Abstract

Convolutional Neural Networks (CNN) are essential in computer vision applications, but their implementation in embedded systems is challenging due to their high memory and computational demands. To address this, techniques such as quantization are employed, allowing the execution of models on embedded hardware, such as FPGAs, that offer energy efficiency and flexibility. Automated development environments such as Vitis AI and FINN from Xilinx facilitate the implementation of CNNs on FPGAs.

This work summarizes CNN models on FPGAs using Vitis AI and FINN for image classification and object detection. Literature and previous work are reviewed, describing architectural differences, model construction and evaluation procedures, and analyzing performance and energy efficiency, highlighting the virtues and limitations of each environment.