Research on Neural Network Compression and FPGA Acceleration Based on Quantization and Pruning

Aiming at the problem that the original AlexNet convolutional neural network suffers from massive parameters and high floating-point computational overhead, which hinders real-time deployment on the Zynq FPGA embedded platform, this paper proposes a joint model compression scheme combining structured channel pruning and KL-divergence-based INT8 quantization. First, L1 regularization is applied for sparse training; 50 % low-contribution convolutional channels are pruned, followed by fine-tuning to restore model accuracy. KL-divergence calibration is then adopted to implement 8-bit integer quantization, which lightweight the model in terms of both computation and storage. Model training and compression are completed with PyTorch, while inference simulation and FPGA deployment are realized via TensorRT and Vivado. Ablation and comparative experiments are carried out on three datasets: COVID-19 chest X-ray images, the German Traffic Sign Recognition Benchmark (GTSRB), and CIFAR-10. Furthermore, dual visualization including Grad-CAM heatmaps and overlaid inference plots is adopted to analyze how lightweight operations affect feature extraction. Experimental results show that the joint pruning-plus-quantization scheme reduces model parameters from 60 M to 28 M, compresses model volume by 88 %, shortens single-image inference latency from 2.40 ms to 0.72 ms with a 3.3-fold speedup, and only brings 0.54 % classification accuracy drop. Visualization demonstrates that the lightweight model can accurately locate target objects with only minor degradation in fine-texture recognition. Under controllable accuracy loss, the proposed scheme significantly cuts hardware resource consumption and provides a practical lightweight solution for portable FPGA-based image-recognition systems in medical and vehicular scenarios.