Optimising Embedded Neural Network Inference in Smart Traps for Fruit Pest Detection via Quantization-Aware Training and FPGA Acceleration

• Lucas C. Freitas UFPel
• Isadora V. Dias UFPel
• Victor R. S. Santos UFPel
• Paulo R. Ferreira UFPel
• Julio C. B. De Mattos UFPel
• Lisane B. De Brisolara UFPel

Resumo

Deploying deep learning models in embedded agricultural systems requires balancing predictive performance with strict hardware constraints such as memory, power, and latency. This work uses quantization techniques, specifically post-training quantization (PTQ) and quantization-aware training (QAT), to optimize convolutional neural networks for real-time pest detection in smart traps. Using the Brevitas framework for quantization and FINN for hardware generation targeting FPGAs, we evaluate a range of weight and activation bit-width configurations. Our results show that QAT significantly outperforms PTQ, particularly in aggressive low-bit scenarios, achieving high accuracy while drastically reducing hardware resource utilization. Among the proposed solutions, one achieves 87.47% accuracy while using less than 10% of the LUTs required by its full-precision counterpart. Comparative analysis with standard models such as ResNet18 and MobileNet further validates the effectiveness of our approach. This study highlights the practicality of QAT-driven quantization for edge Artificial Intelligence applications in agriculture. It paves the way for future work, including power, latency, and throughput profiling, to support large-scale deployment.

Palavras-chave: Training, Deep learning, Analytical models, Quantization (signal), Accuracy, Throughput, Real-time systems, Table lookup, Field programmable gate arrays, Standards, image processing, deep learning, quantization, design space exploration, hardware accelerators