Research Article

Custom Network Quantization Method for Lightweight CNN Acceleration on FPGAs

Table 5

The comparison between the conventional quantized networks and the optimized networks.

NetworkTop-1 accuracy (%)XCZU3EGXCVU13PSize (M)Compression ratio (%)
Inference time (ms)Time reduction (%)Inference time (ms)Time reduction (%)

MobileNetv1-C89.9829.448.043.770.86
MobileNetv3-C93.9053.9714.925.170.17
PPLCNet-C89.5616.354.302.271.12
PPLCNetv2-C93.6147.8113.796.074.35
MobileNetv1-O89.7716.4444.164.3246.273.572.44
MobileNetv3-O93.8418.8964.995.0166.424.772.51
PPLCNet-O89.689.8839.572.6937.442.073.75
PPLCNetv2-O93.4623.0151.876.1055.765.775.64