Affective Abstract Image Classification Based on Convolutional Sparse Autoencoders across Different Domains

To apply unsupervised feature learning to emotional semantic analysis for images in small sample size situations, convolutional sparse autoencoder based self-taught learning for domain adaption is adopted for affective classification of a small amount of labeled abstract images. To visually compare the results of feature learning on different domains, an average gradient criterion based method is further proposed for the sorting of weights learned by sparse autoencoders. Image patches are first randomly collected from a large number of unlabeled images in the source domain and local features are learned using a sparse autoencoder. Then the weight matrices corresponding to different features are sorted according to the minimal average gradient of each matrix in three color channels. Global feature activations of labeled images in the target domain are finally obtained by a convolutional neural network including a pooling layer and sent into a logistic regression model for affective classification. Experimental results show that self-taught learning based domain adaption can provide training data for the application of unsupervised feature learning in target domains with limited samples. Sparse autoencoder based feature learning across different domains can produce better identification effect than low-level visual features in emotional semantic analysis of a limited number of abstract images.