Real-time and Resource-efficient Embedded Computer Vision Via Optimizing Lightweight CNNs for FPGA Acceleration

초록

Deploying deep learning models on resource-constrained edge devices requires efficient hardware acceleration to balance accuracy, latency, and resource utilization. This study optimizes and implements three lightweight convolutional neural networks -MobileNetV2, SqueezeNet, and ResNet18-on field-programmable gate array (FPGA) using transfer learning with the PyTorch framework of Xilinx Vitis AI. Through hardware-aware quantization, model pruning, and FPGA-specific optimizations, we achieve significant reductions in inference latency and resource consumption while maintaining high classification accuracy. To comprehensively evaluate performance, we compare our FPGA implementation against CPU and GPU alternatives. Experimental results show that our optimized models achieve Top-1 accuracies of 91.05%, 89.27%, and 94.73% for MobileNetV2, SqueezeNet, and ResNet18, respectively, on the CIFAR-10 dataset, with inference latencies of 10 ms, 6.28 ms, and 19.95 ms/frame. Detailed FPGA implementation and resource utilization analysis demonstrate the scalability and efficiency of our approach. These findings highlight the potential of FPGA-based deep learning deployment for real-time AI applications in autonomous systems and edge computing.

키워드