An SSD-MobileNet Acceleration Strategy for FPGAs Based on Network Compression and Subgraph Fusion

Over the last decade, various deep neural network models have achieved great success in image recognition and classification tasks. The vast majority of high-performing a huge number parameters often require sacrificing performance accuracy when they are deployed on mobile devices with limited area power consumption. To address this problem, we present an SSD-MobileNet-v1 acceleration method based compression subgraph fusion for Field-Programmable Gate Arrays (FPGAs). Firstly, regularized pruning algorithm sensitivity analysis Filter Pruning via Geometric Median (FPGM) was proposed. Secondly, Quantize Aware Training (QAT)-based full quantization designed. Finally, strategy computing is proposed FPGAs to achieve continuous scheduling Programmable Logic (PL) operators. experimental results show that using can reduce model by factor 11 increase inference speed FPGA platform 9–10. applicable edge be applied real-time monitoring forest fires improve intelligence fire detection.