Learning multi-level representations for affective image recognition

Abstract Images can convey intense affective experiences and affect people on an level. With the prevalence of online pictures videos, evaluating emotions from visual content has attracted considerable attention. Affective image recognition aims to classify conveyed by digital images automatically. The existing studies using manual features or deep networks mainly focus low-level high-level semantic representation without considering all factors. To better understand how are working for tasks, we investigate convolutional visualization them in this work. Our research shows that hierarchical CNN model relies information while ignoring shallow details, which essential evoke emotions. form a more general discriminative representation, propose multi-level hybrid learns integrates semantics representations sentiment classification. In addition, study class imbalance would performance as main category dataset will overwhelm training degenerate networks. Therefore, new loss function is introduced optimize model. Experimental results several datasets show our outperforms various studies. source code publicly available.