A Traffic Sign Recognition Model with Only 140 KB

To design a sign recognition model with low computational complexity and Low parameter quantity, we uses Group Convolution to compress the parameters, and designs extreme block to solve the problem that the number of input channels of Group Convolution must be equal to the number of output channels and that the feature can not be extracted across channels. In this paper, the number of convolution kernels is set according to the number of classifications. Finally, the original 30 MB CifarNet is compressed into a 140 KB classification model. And we tested it on the BelgiumTS Dataset. The experimental test results show that after the model size is compressed to the original 1/220, top1 is not reduced, but it is increased by 87.31%, and top5 is increased by 0.5%. Experiments prove that the compression strategy is effective. And the experiment also explored the relationship between the number of convolution kernels and the number of classifications.