基于典型样本的卷积神经网络技术

Convolutional neural network based on typical samples

针对传统卷积神经网络训练过程中,对于全量样本直接进行特征提取会带有过多非关键区分特征使得训练存在模型过拟合、训练收敛慢等问题,提出一种基于典型样本的卷积神经网络TSBCNN。通过部分典型样本生成强化因子指导修正CNN训练,在特征提取阶段更加注重关键区分特征部分,有目的地降低网络训练过程中对非关键特征的学习,有效提高网络训练效果。大量实验结果表明,TSBCNN较传统CNN网络收敛速度和分类准确率有所提高,在一定程度上有效减少过拟合。

Aiming at the problems such as over-fitting of model,slow convergence of training caused by direct feature extraction of full-size samples,which has abundant non-critical distinguishing features in the process of traditional convolutional neural network training,a CNN model based on typical samples(TSBCNN)was proposed.The enhancement factors were generated by a small amount of typical samples to guide and revise CNN training.In the feature extraction stage,TSBCNN focused more on distinguishing features,reducing the learning of non-critical features in the network training process deliberately,improving the network training effect.Results of experiments show that the convergence speed and classification accuracy of TSBCNN are significantly higher than that of traditional CNN,and over-fitting is effectively reduced partly.