摘要： 实际图像识别场景中样本易受噪声的影响,且大量训练样本的获取比较困难。针对上述问题,提出一种基于混合结构的深度学习网络。选择合适稀疏限制算法运用到自动编码器中,将其构造在混合结构的深度学习网络的底层和高层,解决一般自动编码器无法有效降低 背景噪声影响的问题。在深度学习网络的中间层插入改进的无向高斯伯努利受限玻尔兹曼机,进行无监督学习得到低层输出的抽象表示。混合结构的深度学习网络利用最高的回归层进行有监督的微调和分类。实验结果表明,与传统的浅层算法和经典深度学习网络相比,该混 合结构的深度网络在大、小样本量的训练下均具有较好的抗噪能力。

关键词: 混合结构, 深度网络, 噪声, 稀疏限制, 小样本量

Abstract: Hybrid structure of deep network which achieves low classification error rate when meeting noisy and limited training data is presented since many real world pictures are under noise.Proper sparse regularization algorithm is applied to auto-encoders,which are set on the lower and higher layers and can distinguish background noise more effectively.Middle layer is proposed on the basis of the undirected Gaussian-Bernoulli Restricted Boltzmann Machine(GRBM) to a better abstract representation.Hybrid structure of deep network uses a regression layer on the top for fine-tuning and classification.Experimental results show that hybrid structure of deep network can achieve better performance under noisy background regardless of number of training samples.

Key words: hybrid structure, deep network, noise, sparse limit, small sample amount

中图分类号: 

• TP18