基于跨领域卷积稀疏自动编码器的抽象图像情绪性分类

发布时间：2018-11-27 07:45

【摘要】：为了将无监督特征学习应用于小样本量的图像情绪语义分析,该文采用一种基于卷积稀疏自动编码器进行自学习的领域适应方法对少量有标记抽象图像进行情绪性分类。并且提出了一种采用平均梯度准则对自动编码器所学权重进行排序的方法,用于对基于不同领域的特征学习结果进行直观比较。首先在源领域中的大量无标记图像上随机采集图像子块并利用稀疏自动编码器学习局部特征,然后将对应不同特征的权重矩阵按照每个矩阵在3个色彩通道上的平均梯度中的最小值进行排序。最后采用包含池化层的卷积神经网络提取目标领域有标记图像样本的全局特征响应,并送入逻辑回归模型进行情绪性分类。实验结果表明基于自学习的领域适应可以为无监督特征学习在有限样本目标领域上的应用提供训练数据,而且采用稀疏自动编码器的跨领域特征学习能在有限数量抽象图像情绪语义分析中获得比底层视觉特征更优秀的辨识效果。
[Abstract]:In order to apply unsupervised feature learning to the image emotional semantic analysis of small sample size, a domain adaptation method based on convolution sparse automatic encoder is used to classify a small number of labeled abstract images. An average gradient criterion is proposed to rank the weights of the automatic encoder, which is used to compare the results of feature learning based on different fields. Firstly, random acquisition of image subblocks on a large number of unlabeled images in the source domain is carried out and local features are learned by sparse automatic encoder. Then the weight matrices corresponding to different characteristics are sorted according to the minimum value of the average gradient of each matrix on the three color channels. Finally, a convolution neural network containing pool layer is used to extract the global feature response of labeled image samples in the target domain, and then the logical regression model is used to classify emotion. The experimental results show that domain adaptation based on self-learning can provide training data for the application of unsupervised feature learning in the field of finite sample target. Furthermore, cross-domain feature learning using sparse automatic encoder can obtain better identification results than the underlying visual features in a limited number of abstract image emotion-semantic analysis.
【作者单位】：西北工业大学电子信息学院;郑州轻工业学院计算机与通信工程学院;
【基金】：陕西省科技统筹创新工程重点实验室项目(2013 SZS15-K02)~~
【分类号】：TP391.41

【参考文献】