基于卷积与稀疏编码的半监督学习方法

Semi-supervised Learning Method Based on Convolution and Sparse Coding

卷积神经网络(CNN)在半监督学习中取得了良好的成绩,其在训练阶段既利用有标记样本,也利用无标记样本帮助规范化学习模型。为进一步加强半监督模型的特征学习能力,提高其在图像分类时的性能表现,本文提出一种联合深度半监督卷积神经网络和字典学习的端到端半监督学习方法,称为Semi-supervised Learning based on Sparse Coding and Convolution(SSSConv);该算法框架旨在学习到鉴别性更强的图像特征表示。SSSConv首先利用CNN提取特征,并对所提取特征进行正交投影变换,下一步通过学习其稀疏编码的低维嵌入以得到图像的特征表示,最后据此进行分类。整个模型框架可进行端到端的半监督学习训练,CNN提取特征部分和稀疏编码字典学习部分具有统一的损失函数,目标一致。本文利用共轭梯度下降算法、链式法则和反向传播等算法对目标函数的参数进行优化,将稀疏编码的相关参数约束于流形上,CNN参数既可定义在欧氏空间,也可以进一步定义在正交空间中。基于半监督分类任务的实验结果验证了所提出SSSConv框架的有效性,与现有方法相比具有较强的竞争力。

Convolutional neural network(CNN)has achieved great success in semi-supervised learning.It uses both labelled samples and unlabelled samples in the training stage.Unlabelled samples can help standardize the learning model.To further improve the feature extraction ability of semi-supervised models,this paper proposes an end-to-end semi-supervised learning method combining deep semi-supervised convolutional neural network and sparse coding dictionary learning,called Semi-supervised Learning based on Sparse Coding and Convolution(SSSConv),which aims to learn more discriminative image feature representation and improve the performance of classification tasks.Firstly,the proposed method uses CNN to extract features and performs orthogonal projection transformation on them.Then,learn the corresponding sparse coding and obtain the image representation.Finally,the classifier of the model can classify them.The whole semi-supervised learning process can be regarded as an end-to-end optimization problem.CNN part and sparse coding part have a unified loss function.In this paper,conjugate gradient descent algorithm,chain rule,and backpropagation algorithm are used to optimize the parameters of the objective function.Among them,we restrict the relevant parameters of sparse coding to the manifold,and the CNN parameters can be defined not only in Euclidean space but also in orthogonal space.Experimental results based on semi-supervised classification tasks verify the effectiveness of the proposed SSSConv framework,which is highly competitive with existing methods.