Edge AI Application and Optimization: FPGA-Accelerated Railway Damage Detection and Progressive Quantization Framework

Abstract

The increasing demand for real-time, low-power artificial intelligence (AI) applications has driven the deployment of deep neural networks on resource-constrained edge devices, such as Field-Programmable Gate Arrays (FPGAs). This thesis focuses on developing an efficient railway damage detection system using FPGA-based hardware and proposes a novel quantization framework to address the limitations posed by hardware constraints. Chapter 1 presents the design and implementation of a railway track detection system, which comprises four main components: an enhanced railway damage image dataset, an FPGA integrated with a convolutional neural network (CNN), a host computer for interaction and visualization, and an intelligent vehicle platform. The intelligent vehicle captures real-time images of railway tracks using a gimbal-mounted camera and transmits them to the FPGA for damage detection. The detection results, including track status and precise damage location, are wirelessly sent to the host computer for visualization. Although the system achieves real-time performance and low power consumption, aggressive model quantization is required to fit the FPGA’s resource limitations, which results in a reduction in accuracy. To mitigate this issue, Chapter 2 introduces a progressive quantization framework that incrementally reduces the bit-width of weights and activations while preserving model accuracy. The framework incorporates advanced techniques, including the Straight-Through Estimator (STE), Arctangent Soft Round (ASR), Minimize Discretization Error (MDE), and a filtering mechanism. It is evaluated on six neural network architectures—ResNet18, ResNet20, VGG7, VGG16, MobileNetV2, and ShuffleNetV2—using CIFAR-10 and Tiny ImageNet datasets. Experimental results demonstrate that the proposed approach significantly improves quantized model accuracy, achieving near-full-precision performance in most cases. Future research will focus on integrating the progressive quantization framework into the railway damage detection system, with the goal of balancing resource efficiency and high detection accuracy. This work provides a practical and scalable solution for deploying AI-based monitoring systems on edge platforms, contributing to real-time edge AI applications in critical infrastructure maintenance.