Popis: |
Convolutional neural network (CNN), the network structure of weight sharing, which significantly reduces the complexity of the model and the number of weights, is a research hotspot in the field of speech analysis and image recognition. However, in most cases, insufficient data or small classification differences will result in poor training results: translation invariance and pooling layer. In this paper, the Cross-Entropy overfitting and sample unbalance problems were adjusted by parameter equalization, and the accuracy increased from 51.5% to 87.3%. At the same time, it is verified that in the same datasets, the identification error of the simplified model on the mobile is less than 2%. |