Abstract: In order to bridge the gap between low-level visual features and high-level emotional semantics, and to alleviate the defects inherent in small sample dataset in abstract paintings emotions recognition datasets, a two-layer transfer learning strategy is introduced into traditional convolutional neural networks and a model for recognizing emotions from abstract paintings is proposed using convolutional neural network with a two-layer transfer learning scheme. According to the hierarchical nature of deep features, a large-scale generalized image dataset is used to learn how extract universal low-level image features. Then the relevant domain dataset is utilized to learn how extract specific high-level semantic features. Finally the abstract painting emotion recognition dataset is used to finetune the network. As shown by our extensive experimental validation on MART datasets, the proposal outperforms current methods when recognizing emotions from abstract paintings.

Key words: emotion recognition, deep learning, transfer learning, convolutional neural network, abstract paintings