自适应深度卷积神经网络模型构建方法

发布时间：2018-06-25 20:52

本文选题：深度卷积神经网络 + 自适应模型构建　；参考：《北京邮电大学学报》2017年04期

【摘要】：针对传统卷积神经网络(CNN)模型构建过度依赖经验知识、参数多、训练难度大等缺点,同时鉴于复杂多类问题的CNN模型构建策略的重要价值,提出一种自适应深度CNN模型构建方法.首先,将初始网络模型的卷积层和池化层设置为仅含一幅特征图;然后,以网络收敛速度为评价指标,对网络进行全局扩展,全局扩展后,根据交叉验证样本识别率控制网络展开局部扩展,直到识别率达到预设期望值后停止局部网络学习;最后,针对新增训练样本,通过拓展新支路实现网络结构的自适应增量学习.通过图像识别实验验证了所提算法在网络训练时间和识别效果上的优越性.
[Abstract]:In view of the shortcomings of the traditional convolution neural network (CNN) model, which is overly dependent on experience knowledge, with many parameters and difficulty in training, an adaptive depth CNN model construction method is proposed in view of the important value of the CNN model construction strategy for complex and multi class problems. First, the convolution layer and the pool layer of the initial network model are set to only one piece of the convolution layer and the pool layer. Then, with the network convergence speed as the evaluation index, the network is extended globally. After the global expansion, the local extension is expanded according to the cross validation sample recognition rate control network, and the local network learning is stopped after the recognition rate reaches the expected value. Finally, the new training route is extended to realize the network structure by expanding the new branch road. Self adaptive incremental learning is used to verify the superiority of the proposed algorithm in network training time and recognition effect through image recognition experiments.
【作者单位】：山东理工大学电气与电子工程学院;哈尔滨工程大学自动化学院;
【基金】：山东省自然科研基金项目(ZR2015FL029,ZR2016FL14) 国家自然科学基金项目(61601266) 中国博士后科学基金项目(2017M612306)
【分类号】：TP183;TP391.41

【相似文献】