鲁棒可预测判别字典学习人脸识别方法

Robust discriminative K-SVD network for face recognition

提出一种可预测判别K-SVD网络模型(DKSVDN)并用于人脸识别问题。该模型构造了一种新颖的字典结构,包含类别标签字典和描述字典,以兼顾判别和重构性能。相应的稀疏编码向量由标签编码向量和描述编码向量组成。针对样本稀疏编码时间效率低的问题,利用预测神经网络与判别字典学习模型协同训练的方法来加速预测稀疏编码。此外,针对DKSVDN还特别引入一种拟梦境的训练方法用于提升模型在训练集多样性不足时的鲁棒性。通过在主流人脸数据集上的对比实验证明了该模型的优良性能。

This paper presented a novel discriminative K-SVD network(DKSVDN)for face recognition.It embedded discriminative information into traditional K-SVD algorithm by special design of dictionary as well as sparse representation coefficients on the dictionary.The dictionary consisted of label specific atoms and descriptive atoms,while sparse codes contained one-hot label vectors and descriptive codes.In addition,as sparse representation algorithms were time-consuming,DKSVDN attached a co-trained feed-forward neural network to discriminative dictionary learning model to predict sparse codes.More-over,with generative module in DKSVDN,this method also designed a new dreaming training phase to improve the robustness of DKSVDN for unknown pattern in known class.The experiment results on public face image datasets verify the effectiveness of this method.